We think hardware and software have to be built together. That's how tech ends up feeling obvious, and good looking.
Your campus. Your schedule. Your people.
Now in testing on iPhone and Android.
Concept computers designed around focus.
A concept: nothing is manufactured yet. Buying opens at release, expected in 2030.
Sign in once, and Shiba, Cadence and Momentum all know it's you.
Three apps, one account, one platform. They're made to work together.
Across every product we're building an AI layer that's actually useful. It reads your context, uses tools carefully, shows its work, and you stay in charge.
Intelligence that does the work.
Parable 1 reasoning, Agent Mode, effort tiers, code artifacts, images, voice and memory.
Available now.
School life, in one place.
Feed, reels, calendar, notes, channels, DMs and verified school registration.
In testing on iPhone and Android.
Visual momentum, nightly.
A glowing streak matrix, a 30-second evening check-in, and private on-device storage.
Available now.
Computers designed around focus.
Two desktops, Aero™ and Aero™ Neo, and two AeroBook™ laptops, Nano and Pro. They go from an 8 GB fanless N100 laptop up to a twenty-four-core desktop, all running our own Linux with the Glass UX. The AI in it is there to help: find a document, sum up a project, show you what's next, and ask before it changes anything important.
A concept. Nothing is manufactured yet.
Send anything, anywhere.
Files between your own devices in a browser tab, and between people from a Mac or an iPhone, encrypted before anything leaves the machine.
In development.
Our SaaS starter, as a zip.
Next.js 15 with authentication, Stripe billing, workspaces and a dashboard already wired together.
A download, $199 USD, paid once, with a code.
One sign-in. Everywhere.
Identity, plan, privacy and OAuth: the seam between every product.
Free with every product.
Shared software foundations.
Shared auth, billing, rate limits, storage, AI gateway and security policies. Every product inherits it for free.
Runs under every Noctic app.
They overlap on purpose. Which one you want depends on what you're trying to do this afternoon, not on which is the most capable.
| Shiba | You have a problem to think through, something to write, code to work on, or a question with more than one right answer. Start here. Everything else runs on the same reasoning. |
|---|---|
| Cadence | You're at a school and need the day in one place: the feed, the calendar, notes, channels and messages, with verified registration so you know who you're talking to. |
| Momentum | You want to keep at something daily and see the evidence. A thirty-second evening check-in, stored on your own device. |
Why build several products and not one? Because the hard parts are the same in all of them. Every app sits on the same foundations, so when one gets better they all do.
Sign in once. Your identity, plan and privacy choices follow you into every product.
The same reasoning model behind Shiba answers in Cadence too.
The same privacy policy, the same safety work and the same limits on what an app may ask for.
Export everything, or delete the account and its content across every product, from Settings.
No. Each one works on its own and none of them bugs you about the others. They get better together because they can read each other's context, not because we lock stuff until you sign up twice.
Yes, on every product. The full breakdown of what each plan includes is on the pricing page.
Not unless you turn it on. It's off by default in Settings. Heads up: right now that switch is saved on the device you flipped it on and doesn't reach our server yet. So treat it as telling us what you want, not as something we enforce, until this page says otherwise.
You need to be at least 13 for a Noctic Account. What we do differently for younger and school-age users is in children and teens.
Yes. Continue with Noctic is open to apps that aren't ours, and the tools to wire it up are in the developer section.
The SaaS starter we build on, as a zip. Next.js 15, authentication, billing, multi-tenant workspaces and a dashboard, all wired together, typed end to end, and unbranded so you can ship it as your own.
Not a template you pull apart. A working product with your name missing from it.
A download, $199 USD, paid once.
JWT sessions in httpOnly cookies with server-side session records, so revocation is instant. Email and password, Google and GitHub over PKCE, protected-route middleware and role-based guards.
Stripe Checkout and the customer portal, webhooks that verify signatures and survive being delivered twice, tables for subscriptions and one-off payments, and plan entitlements.
A Drizzle schema for users, OAuth accounts, sessions, organisations, members, invitations, subscriptions and payments, with the SQL migrations already generated.
A token-driven component library: buttons, cards, inputs, tables, modals, dropdowns, toasts, avatars and badges. Dark and light, responsive, no framework to learn on top of Tailwind.
Collapsible sidebar, workspace switcher, user menu, theme switching with no flash on load, and the settings, team and billing pages already written against the schema.
npm run typecheck and npm run build both pass on a fresh unzip. Three
environment variables boot it; everything else stays disabled until you add its keys.
Next.js 15 with the App Router, TypeScript, Tailwind CSS, Drizzle ORM, PostgreSQL and Stripe. The zip is
the source tree, not a node_modules dump. Unzip it and run npm install.
No checkout here and no card form. You send the money from your own bank, we send you a code, and the code opens the download.
$199 USD, once
Send it to
nicholassteiche@icloud.com
Security question, if your bank asks for one
What kit is this
Answer
kickstart20
Interac e-Transfer to nicholassteiche@icloud.com. If your bank asks for a security question, use the one above: What kit is this, answered kickstart20, lower case and no space.
Write the address on your Noctic Account in the transfer's message field. That's where the code is sent, and the account it belongs to.
Once the transfer lands in the bank we email you a code of your own. It looks like
KS-A7QF-2M4X-9KDP-3TWE, and no one else is ever given it.
Enter the code below and the zip downloads. The code is used up in the act, so save the file instead of the code.
Paying from outside Canada, or in a currency your bank won't e-Transfer? Write to support and we'll sort out another way to take the $199.
Sign in first. The code is issued to your account. Entering it uses it up and opens a private link that lasts fifteen minutes, so download the zip and keep it. A used code is used for good, including for anyone it gets forwarded to.
Your code is spent
Kick Start is yours. Save the zip somewhere you'll find it again, then unzip it, run
npm install, copy .env.example to .env.local, and the README
takes it from there.
| The whole source | Every file, readable and editable. Nothing minified, nothing hidden, nothing saved for a pricier tier. |
|---|---|
| Unlimited projects | Build as many products on it as you want, for you or for clients, and charge what you like. One purchase covers all of them. |
| Not for resale as a kit | Ship what you build with it. Don't repackage the kit itself, changed or unchanged, as a starter for other people to buy or download. |
| No subscription | One payment. There's no renewal and no seat count. What you download is yours to keep and to build on for as long as you like. |
| Support is best effort | Ask us anything at contact. This is a zip, not a managed service, so treat any answer as a favour instead of an entitlement. |
| No refunds after the code | The code is the product, and it can't be returned once it's sent. If something in the kit is actually broken, tell us and we'll fix it or refund you. |
Because a card checkout would mean running a whole payment setup for one product. An e-Transfer costs you nothing, clears the same day, and your card details never come near us.
Usually a few hours, one business day at worst. A real person reads the transfers, so if you pay overnight you'll hear back in the morning.
Send the Canadian equivalent of $199 USD on the day you pay. If the rate's off a little either way that's fine. We're not chasing anyone over a dollar.
Check it character by character. Codes never use the letters I or O or the digits 0 or 1, so those are usually the typo. Spaces and lower case are fine. If it still fails, reply to the email the code came in.
No, and that's on purpose. A spent code is dead for everyone, so a forwarded one is worth nothing. For the first day after you use it, we'll still give you a fresh link if a download drops halfway. After that, ask us and we'll open it again.
No. Kick Start is sold and delivered entirely outside the account system. Nothing about this purchase touches your Noctic Account if you have one.
Dependencies, which npm install fetches; a database, which you point
DATABASE_URL at; and your Stripe and OAuth keys. The README lists every variable and
what it switches on.
You get the build from that day, and the code won't open a newer one. If you want the current kit again, write to us and we'll send a fresh code. There's no changelog feed and no way to upgrade a project you've already changed.
Encrypted before it leaves your machine. Usync moves files between your own devices in a browser tab. Usync Air moves them between people, from a Mac or an iPhone, and nothing in the middle can ever read them.
AirDrop only works when both machines are in the same room. Email attachments get read by every server they pass through. This is the third option.
In development.
Usync
A browser tab on every machine you own. Drop files, a folder, some text or a link into one and it shows up on another. No cable, no juggling accounts, no emailing yourself.
Names your devices, remembers the ones you use, and works just as well from a phone browser as a desktop. Comes with a Noctic Account.
Usync Air
Native apps for macOS and iPhone that send to approved contacts. Watched folders, team spaces, and a relay that holds ciphertext for as long as it takes the other side to collect it.
$10 a month. Universal on Apple silicon and Intel, iOS 16 or later.
Files that show up on a machine by themselves are how machines get wrecked. So a security code lets someone reach you. It never lets them write to you. That needs your OK, once, per person.
Sixteen characters, generated for you and never chosen. Share it like a phone number. Roll it the moment it leaks and the old one dies instantly.
That pair finds a person and asks to connect. It moves zero bytes. Wrong guesses are rate limited and look identical to misses, so codes can't be fished out one attempt at a time.
Both sides agree before the database permits a single transfer. Auto-download then runs only for approved contacts, and only under the ceiling you set.
Everything is sealed on your device before it's uploaded. What passes through is ciphertext and a checksum, and it's deleted the moment it's collected.
Five steps, and none of them touch a server that could read anything even if it tried.
| 1 · Sealed | The sender's device generates a one-time key and encrypts the file with AES-256-GCM. The key exists for this one transfer and is never reused. |
|---|---|
| 2 · Wrapped | That key is wrapped for the recipient using X25519 key agreement, so only their device can unwrap it. Private keys live in the Keychain and never leave the machine they were made on. |
| 3 · Relayed | The ciphertext goes to the relay, which holds no key and can decrypt nothing. It's a corridor, not a filing cabinet. |
| 4 · Verified | The recipient downloads, unwraps, decrypts and recomputes the SHA-256. If it doesn't match, it's thrown away instead of saved. Half a file isn't a file. |
| 5 · Wiped | Once collected, the object is deleted from the relay. Nothing accumulates that could later leak. |
Point Usync Air at a folder. Save a file into it and the file is already on the other machine. No button, no upload dialog, no remembering to send it.
Chosen per contact. People you trust land in a folder. Everyone else waits in an inbox until you say so, however much they send.
Group your people. Drop a file once and it reaches everyone approved in the space.
Every transfer carries a SHA-256 checksum, recomputed after decryption. A mismatch is refused, not opened.
Nothing needs to be on the same Wi-Fi, in the same building or on the same continent. If both machines can reach the internet, they can reach each other.
The same Noctic Account that carries Shiba and Cadence. Sign in once and the apps know who you are.
Some settings are up to you. These aren't, because the whole point of auto-download is that you're not there to catch anything.
| Executables never auto-save | Whatever auto-download says, anything on the refused list waits for you to look at it. The list is in the app, not a server setting somebody else can edit. |
|---|---|
| Quarantine is kept | Files written to disk keep the flag macOS puts on downloads, so the operating system still asks its own questions before anything runs. |
| Your size cap holds | Anything over the ceiling you set waits in the inbox instead of filling a disk while you sleep. |
| Approval is per person | Approving one contact approves that contact. There's no setting that opens you to everybody, because we didn't build one. |
| Keys stay on the device | Private keys are generated in and never leave the Keychain. We can't read your transfers, and there's no key escrow to be compelled for. |
Included with a Noctic Account
$10 USD / month
Downloads are tied to your account and plan, so a build can't be passed around. Stripe handles payment and card details never reach us, same as everything else on the pricing page.
The apps are real and they build. What's not real yet is everything behind them, and it'd be easy to write this page like it was.
macOS compiles in release and ships as a disk image. iOS runs. The encryption, the contact model and the safety rules above are actually written, not just sketched.
The database schema and the functions that turn a Noctic sign-in into a session are written but not live yet. Every call fails with a clear message so nothing pretends to work.
Until it's signed and notarised, macOS will refuse it on any machine but the one that built it. That's the last step before anyone else can run this.
A limit of the current storage plan, not the design. Resumable chunked upload is the next piece of work and lifts it.
No. And that's not a promise, it's how it's built. The key that opens a file is made on your device and wrapped so only the recipient's device can unwrap it. All we ever hold is ciphertext we have no key for, briefly, and then nothing.
The private key goes with it, which means transfers sent to that device and not yet collected can't be opened by anything. Set up a new device and ask the sender to send again. There's no recovery key, because a recovery key is a key somebody else can be made to hand over.
Ask to connect. That's the whole list. Until you approve them the database refuses every transfer, and you can roll the code from the app, which kills the old one immediately.
AirDrop needs both machines nearby and both to be Apple. This needs an internet connection and an approved contact. The overlap is smaller than it sounds.
No. The browser half stands alone and covers sending things to your own devices, which is most of what most people want. Air is for sending to other people.
Not yet, and not soon. The browser half runs anywhere; the native apps are macOS and iOS because that's where the Keychain work is done and where we can test properly.
When the backend is deployed and the installer is signed. Both are on the list above, and neither has a date, because a date we can't keep is worse than no date at all.
New to Linux? No problem. AXESS OS is easy to learn and easy to live with. If you're thinking about moving from Windows or a Mac, here are the questions people ask first.
Yes. It's Linux underneath and Glass UI on top. Settings are named after what they do. No daemons, no mount points, no acronym you meet once a year. The terminal is there if you want it, and you never have to open it to do ordinary computer things.
See Glass, the design languageNormal Linux software runs, AppImage packages work as is, and a small set of our own apps comes with the basics installed. All four machines use the same Intel parts today, so what you install on one works on every Aero™ we ship now. Noctic Silicon is RISC-V, so when we move to our own chip, the system and its packages get rebuilt for it.
Because then every port is the same port: power, displays, storage, audio and networking through one connector, and any cable in any socket. The space an HDMI socket would take becomes battery on a laptop and airflow on a desktop. On day one you'll need an adapter for something. One hub on your desk covers it.
Nothing you haven't turned on. No keystroke logging, no screen watching, no “improve the product” toggle that shows up already on, and saying no never switches off something you paid for. Every machine works fully signed out.
How privacy worksNot yet. The AeroBook™ Nano is designed and its $499 is settled, but it hasn't been manufactured, so there's nothing to ship and we aren't taking orders. The other three are concepts with indicative pricing. Reserving a spot is free and commits you to nothing.
Reserve a spot
The same build of AXESS OS runs on all four machines, desk or bag. Every control is drawn in Glass UI, with no borrowed system dialogue anywhere, and the terminal is still there for anyone who wants it.
See Glass
Standard AppImage packages work, alongside a small first-party library and useful essentials installed from the start. It's still Linux, presented better.
The desktops come without a display, keyboard or mouse: anything that speaks USB-C or Bluetooth works and pairs from the first screen you see. Up to four displays on the Aero™, one of them at 8K.
Every machine works fully signed out. A Noctic Account syncs your settings between machines if you want that, and nothing is held back if you don’t.
How privacy works
The Aero™ pairs Intel’s Core Ultra 9 285 (eight performance cores and sixteen efficiency cores, up to 5.6 GHz) with 64 GB of memory, under one large fan that barely has to turn.
A vapour chamber (a sealed copper plate with a little liquid inside) boils where the chip is hot and condenses at the cool edges, over and over, with no pump and nothing to wear out. The AeroBook™ Nano needs no fan at all.
Nothing only charges and nothing only carries data. USB-C for power, displays, storage, audio and networking. Up to 80 Gb/s per port on the AeroBook™ Pro.
Soldered next to the processor instead of in slots: shorter distances, lower latency, and no upgrading later. Buy the size you'll still want in four years.
Up to 14 hours on the AeroBook™ Nano and up to 16 on the Pro, whose 100 Wh pack is more than twice the Nano’s, because a machine that can draw power will.
The family is designed around our own N1 chips: N1x, N1 Bionic and N1 Pro. Until those exist it runs Intel Core Ultra, and an N100 in the Nano. Both are Intel, so it's the same AXESS OS, same apps, same drivers.
Read the planThe Nano’s $499 is final. The Aero™ Neo from $1,499, the Aero™ from $2,499 and the AeroBook™ Pro from $3,499 are indicative until release in 2030, and every page says so next to the number.
A vapour chamber with nothing to wear out, memory sized for the next four years, and one architecture for the whole family. Moving to our own chips one day is a software non-event.
The AeroBook™ Nano is partly made for schools. It's light, silent, cheap to replace, and gets the same AXESS OS updates on the same day as the Pro.
Noctic for EducationThe warranty is intended at two years, parts and labour, on every machine in the family.
Intel N100 · 8 GB
Light, silent, and $499. 12.5″, 1.02 kg, no fan, up to 14 h.
$499 · final
Core Ultra 9 285H · 32 GB
Sixteen cores that travel. 16.2″ 144 Hz, 2.14 kg, up to 16 h.
From $3,499 · indicative
Core Ultra 9 285H · 32 GB
Desktop-class, fifteen centimetres square. Runs from a USB-C charger.
From $1,499 · indicative
Core Ultra 9 285 · 64 GB
Twenty-four cores, one fan barely turning, room for four displays.
From $2,499 · indicative
Not sure? Answer seven questions or compare the full specifications.
Seven questions, then a straight answer, including “none of these”.
Answer themThe whole plan for building our own RISC-V chips, including the ways it could fail.
Read the planThe first hundred, for the three machines that aren't for sale yet. It costs nothing.
Reserve oneWhat every control in AXESS OS is drawn with, and the rules behind it.
See GlassAnything this page didn't answer, including whether the Nano is enough for you.
Get in touchThe AeroBook™ Nano is partly made for classrooms. Here's the rest of that story.
Noctic for EducationEvery decision about this family, with dates, including the ones we took back.
Read the changelogPlanned numbers for machines that mostly aren't built yet. Any of them can change.
| Aero™ Neo | Aero™ | AeroBook™ Nano | AeroBook™ Pro | |
|---|---|---|---|---|
| Processor | Intel Core Ultra 9 285H | Intel Core Ultra 9 285 | Intel Processor N100 | Intel Core Ultra 9 285H |
| Platform | Core Ultra 200H series | Core Ultra 200S series | Alder Lake-N | Core Ultra 200H series |
| CPU cores | 16 (6P + 8E + 2LP-E) | 24 (8P + 16E) | 4 (4E) | 16 (6P + 8E + 2LP-E) |
| Threads | 16 | 24 | 4 | 16 |
| Max turbo | 5.4 GHz | 5.6 GHz | 3.4 GHz | 5.4 GHz |
| L3 cache | 24 MB | 36 MB | 6 MB | 24 MB |
| Graphics | Arc 140T · 8 Xe cores | Intel Graphics · 4 Xe cores | UHD Graphics · 24 EU | Arc 140T · 8 Xe cores |
| Neural engine | NPU 3 · 13 TOPS | AI Boost NPU · 13 TOPS | None | NPU 3 · 13 TOPS |
| Memory | 32 GB soldered | 64 GB soldered | 8 GB soldered | 32 GB soldered |
| Memory speed | LPDDR5X-8400 | DDR5-5600 | LPDDR5-4800 | LPDDR5X-8400 |
| Upgradeable | No. Memory and storage are fixed when the machine is made, on all four. | |||
Intel, Core Ultra and Arc are trademarks of Intel Corporation. Noctic has no partnership, endorsement or supply deal with Intel. These are just the parts machines like these would use if they were built today, picked off a public price list like anyone else’s.
Our own RISC-V chips: N1x, N1 Bionic and N1 Pro. Three chips, one architecture, so one build of AXESS OS runs on all of them, each a step up from the last. None of them has been made yet. Every number below is a target in a spec, not something measured off a chip.
| Aero™ Neo | Aero™ | AeroBook™ Nano | AeroBook™ Pro | |
|---|---|---|---|---|
| Chip | N1 Bionic | N1 Pro | None. Stays on Intel N-series | N1 Pro |
| CPU cores | 8 (4P + 4E) | 24 (8P + 16E) | None | 24 (8P + 16E) |
| Threads | 8 | 24 | None | 24 |
| Max turbo | 4.2 GHz | 5.7 GHz | None | 5.7 GHz |
| L3 cache | 16 MB | 36 MB | None | 36 MB |
| GPU cores | 10 | 24 | None | 24 |
| Neural engine | 16-core · 12.3 TOPS | 32-core · 48 TOPS | None | 32-core · 48 TOPS |
| Memory | 16 GB LPDDR5X | 16 GB | None | 16 GB |
| Memory bandwidth | 136.5 GB/s | 205 GB/s | None | 205 GB/s |
| Status | Designed and running in simulation. Not manufactured, no date. | Specification only. Not designed, not manufactured, no date. | Designed and priced. Not manufactured, no date. | Specification only. |
| Aero™ Neo | Aero™ | AeroBook™ Nano | AeroBook™ Pro | |
|---|---|---|---|---|
| Display | None. Bring your own over USB-C | 12.5-inch, 1920 × 1200 IPS, matte, 60 Hz, 400 nits | 16.2-inch 8K Ultra Retina XDR, 144 Hz, 1600 nits peak | |
| Colour | Whatever your display shows | sRGB, 8-bit | P3 wide gamut, 10-bit, factory calibrated | |
| Touch and unlock | Password, PIN, or a paired AeroBook™ | No touch · password or PIN | Multi-touch · face unlock | |
| Keyboard | Not included. Any USB-C or Bluetooth keyboard works | 1.3 mm travel, not backlit, with the Aero™ key | Backlit, 1.5 mm travel, with the Aero™ key | |
| Camera | None | 1080p, in the top bezel | 12 MP, in the Dynamic Notch | |
| Aero™ Neo | Aero™ | AeroBook™ Nano | AeroBook™ Pro | |
|---|---|---|---|---|
| Storage | 1 TB NVMe · 2 or 4 TB to order | 2 TB PCIe 5.0 NVMe · 4 or 8 TB to order | 64 GB eMMC or 500 GB NVMe | 2 TB PCIe 5.0 NVMe · 4 or 8 TB to order |
| Sequential read | Around 7 GB/s | Around 14 GB/s | 0.3 GB/s or 2.2 GB/s | 7.4 GB/s |
| Power | USB-C, 140 W adapter | USB-C, 240 W adapter | 42 Wh battery · 45 W USB-C | 100 Wh battery · 96 W USB-C |
| Battery life | No battery | Up to 14 h | Up to 16 h | |
| Cooling | Vapour chamber and one quiet fan | Vapour chamber and one 140 mm fan | Vapour chamber, no fan | Vapour chamber and two fans |
| Size | 15 × 15 × 6 cm | 16 × 16 × 27 cm | 12.5-inch · 15.4 mm thick | 16.2-inch · 16.8 mm thick |
| Weight | 1.5 kg | 4.1 kg | 1.02 kg | 2.14 kg |
| Body | Aluminium unibody | Aluminium | Anodised aluminium | Titanium-and-aluminium unibody |
| Aero™ Neo | Aero™ | AeroBook™ Nano | AeroBook™ Pro | |
|---|---|---|---|---|
| USB-C ports | 4 | 6 | 2 | 4 |
| Per-port bandwidth | USB4 · 40 Gb/s | USB4 · 40 Gb/s | USB 3.2 Gen 2 · 10 Gb/s | USB4 v2 · 80 Gb/s |
| External displays | Three, up to 6K | Four, one at 8K | One 4K | Three 6K or one 8K |
| Power | USB-C, 140 W | USB-C, 240 W | USB-C, 45 W | USB-C, 96 W |
| Wireless | Wi-Fi 7 and Bluetooth 6.0 | Wi-Fi 7 and Bluetooth 6.0 | Wi-Fi 6E and Bluetooth 5.3 | Wi-Fi 7 and Bluetooth 6.0 |
| Aero™ Neo | Aero™ | AeroBook™ Nano | AeroBook™ Pro | |
|---|---|---|---|---|
| Speakers | One, for alerts | Two | Six with dual subwoofers | |
| Microphones | None | Two | Four, studio array | |
| System | AXESS OS: Linux-based, Glass UI, AppImage support | |||
| Telemetry | None unless you switch it on | |||
| Warranty | Intended at two years, parts and labour | |||
| From | $1,499, indicative | $2,499, indicative | $499, final | $3,499, indicative |
| Its own page | Aero™ Neo | Aero™ | AeroBook™ Nano | AeroBook™ Pro |
Concept status: the Aero™, Aero™ Neo and AeroBook™ Pro are concepts. Nothing has been manufactured, there's no manufacturing partner and no release date. Their prices in the builder are indicative, and buying is locked until release, which we don't expect before 2030. The processor choice, memory, storage, display, battery, cooling, body and every other hardware and software detail on this page is planned, not final, and may change or be abandoned. Intel, Core Ultra and Arc are trademarks of Intel Corporation; Noctic has no partnership with or endorsement from Intel. The N1x and N1 Pro are specifications for silicon that doesn't exist yet, and N1 Bionic is designed but not made. See Noctic Silicon.
The cheapest machine we could design without it being a bad one. Intel N100, 8 GB of memory, no fan, and the same AXESS OS as the other three.
Designed and priced. Not yet manufactured.
$499, settled. Designed, but not manufactured, so there's no date and no order to place yet.
Every other machine in the lineup is a premium machine. This one isn't, on purpose. It's the only real way to put AXESS OS in front of a student, a classroom, or anyone who just doesn't want to spend a lot on a laptop.
The AeroBook™ Nano is built around a processor that costs a fraction of the others, a body made of anodised aluminium instead of titanium, and a screen picked for being good, not remarkable. What it doesn't cut is the software. It runs the same build of AXESS OS, gets the same updates on the same day, and collects the same amount of data about you as the Pro: none.
Browsing, email, documents, spreadsheets that a person made instead of a bank, calls, note-taking, streaming, homework, coursework, and being carried around all day without a charger.
Video editing, 3D, gaming past what a browser can do, big compiles, virtual machines, or forty tabs at once. It'll try all of these and you'll notice it struggling.
Four cores, four threads and 8 GB. About twenty-five tabs and a document is a comfortable day. Past that it starts swapping and you start waiting. Better we tell you now than you find out in week three.
Six watts of base power means no fan, no vents whirring in a quiet room, and nothing inside that can fail mechanically. Being fanless is a result of the chip, not some feat of engineering. A thin vapour chamber spreads the heat across the body instead. It's a sealed plate with a little liquid inside that boils off the hot chip and condenses at the cool edges, so it moves heat like a tiny fridge with no pump.
Glass UI, the same settings, the same apps, the same terminal if you want it. We don't hold anything back on the cheap machine to make the expensive ones look better.
No neural engine, so on-device AI falls back to the CPU and runs slow or not at all. No face unlock. No touch. A 60 Hz screen. That's what the price costs you, and it's real.
A four-core Alder Lake-N chip that draws six watts and costs almost nothing. It's the most boring processor in the lineup and the easiest one to defend.
The N100 has four efficiency cores, no performance cores, no hyper-threading and no neural engine. It turbos to 3.4 GHz, shares 6 MB of cache across all four cores, and runs a small built-in GPU that's enough for a 1200p desktop, video, and whatever graphics a web page needs. It's the same chip inside tons of mini PCs and thin clients, and that's exactly why we'd use it. It's cheap, it's everywhere, Linux supports it well, and nobody has to invent anything to make it work.
| Processor | Intel® Processor N100 (Alder Lake-N) |
|---|---|
| Cores / threads | 4 efficiency cores / 4 threads (no hyper-threading) |
| Frequency | Up to 3.4 GHz burst |
| Cache | 6 MB L3, shared |
| Base power | 6 W |
| Graphics | Intel® UHD Graphics: 24 execution units, up to 750 MHz |
| Neural engine | None. On-device AI runs on the CPU, when it runs at all. |
| Memory | 8 GB LPDDR5-4800, single channel, soldered to the board |
| Upgradeable | No. Nothing socketed, nothing removable. |
The Nano’s 8 GB is soldered on the board next to the chip, same as every Aero™ and AeroBook™. It's slower than the LPDDR5X in the AeroBook™ Pro, you can't upgrade it, and that's part of hitting the price.
The N100 has one memory channel, so it gets about a third of the AeroBook™ Pro's bandwidth. You feel it in graphics and anything that moves a lot of data at once. You won't feel it writing an essay.
The N1 chips are planned for the machines that need them. The Nano is about being cheap and available now, so for a machine like this the smart move is to keep buying the cheapest decent chip out there. The whole silicon story is on the Noctic Silicon page.
It tops out low, the memory is soldered, and there's only one configuration. If you already know 8 GB won't cut it, look at the Aero™ Neo or the AeroBook™ Pro. That's a straight answer, not an upsell.
The only choice you make when you buy a Nano, and it matters more than it looks.
| Nano 64 GB | Nano 500 GB | |
|---|---|---|
| Medium | 64 GB eMMC 5.1, soldered | 500 GB NVMe SSD, M.2 2242 |
| Sequential read | Around 300 MB/s | Around 2.2 GB/s |
| Free after AXESS OS | Roughly 48 GB | Roughly 484 GB |
| Made for | Cloud-first work, school fleets, a second machine | Keeping your files on the machine |
| Everything else | Identical. Same chip, same memory, same body, same screen, same system. | |
A school account, a cloud drive, a home server. The machine is just a window into them, and it costs as little as it can.
Photos, a music library, coursework going back years, anything you want to open on a train with no signal.
Booting up and everyday stuff feel about the same. Copying a few gigabytes doesn't. We're not going to act like they're equal.
The eMMC is soldered. The NVMe module isn't, but we won't support you opening the machine.
Anodised aluminium, not the titanium-and-aluminium unibody of the AeroBook™ Pro. On a machine this size titanium mostly shows up as a cost on a spreadsheet. The lid is stiff, the hinge opens with one hand, and the whole thing weighs a bit over a kilogram.
| Display | 12.5-inch, 1920 × 1200, IPS, matte, 400 nits, 60 Hz, sRGB |
|---|---|
| Touch | No. The AeroBook™ Pro has it; this one doesn't. |
| Camera | 1080p, in the top bezel. No notch, no face unlock. |
| Unlock | Password or PIN |
| Keyboard | 1.3 mm travel, not backlit, with the Aero™ key |
| Speakers | Two, bottom-firing |
| Microphones | Two |
| Ports | Two USB-C, USB 3.2 Gen 2 at 10 Gb/s, charge or data on either |
| External display | One, up to 4K at 60 Hz |
| Wireless | Wi-Fi 6E and Bluetooth 5.3 |
| Battery | 42 Wh, up to 14 hours, 45 W USB-C charging |
| Cooling | Vapour chamber, no fan. No vents, no moving parts. |
| Weight and thickness | 1.02 kg · 15.4 mm |
| Body | Anodised aluminium |
USB-C only, like the rest of the lineup. On the Nano that means USB 3.2 instead of USB4, because the N100 doesn't offer Thunderbolt. One cable still carries power, a display, storage and audio. If you need HDMI or USB-A, a hub gives you both.
Planned figures for an unbuilt machine. Every one of them may change.
| Model | AeroBook™ Nano |
|---|---|
| Processor | Intel® Processor N100 |
| Cores / threads | 4 / 4 |
| Max frequency | 3.4 GHz |
| Cache | 6 MB L3 |
| Base power | 6 W |
| Graphics | Intel® UHD Graphics, 24 EU |
| Neural engine | None |
| Memory | 8 GB LPDDR5-4800, single channel, soldered |
| Storage | 64 GB eMMC or 500 GB NVMe |
| Display | 12.5″ 1920 × 1200 IPS, 400 nits, 60 Hz |
| Ports | 2 × USB-C, 10 Gb/s |
| Wireless | Wi-Fi 6E, Bluetooth 5.3 |
| Battery | 42 Wh · up to 14 h |
| Cooling | Vapour chamber, fanless |
| Weight | 1.02 kg |
| Thickness | 15.4 mm |
| Body | Anodised aluminium |
| System | AXESS OS: Linux-based, Glass UI, AppImage support |
| Telemetry | None unless you switch it on |
| Warranty | Intended at two years, parts and labour |
| Intended successor silicon | None planned. See Noctic Silicon |
| Aero™ Neo | Aero™ | AeroBook™ Nano | AeroBook™ Pro | |
|---|---|---|---|---|
| Type | Desktop | Desktop | Laptop | Laptop |
| Processor | Core Ultra 9 285H | Core Ultra 9 285 | N100 | Core Ultra 9 285H |
| Threads | 16 | 24 | 4 | 16 |
| Memory | 32 GB | 64 GB | 8 GB | 32 GB |
| Screen | Bring your own | Bring your own | 12.5″ 60 Hz | 16.2″ 144 Hz |
| Weight | 1.5 kg | 4.1 kg | 1.02 kg | 2.14 kg |
| From | $1,499 | $2,499 | $499 | $3,499 |
| Read more | Aero™ Neo | Aero™ | You're here | AeroBook™ Pro |
A browser, a document, mail, music and a video call all at once? Yes, easily. A big photo library, a virtual machine or forty tabs? No, and no configuration of this machine changes that. That's what the AeroBook™ Pro and the Aero™ Neo are for.
Because the moment you move up, the price moves with it, and then it's just a more expensive laptop with the same screen. The Nano is only worth building if it's actually cheap.
64 GB if your files live in an account somewhere and the machine is a window onto them. 500 GB if you keep things on the machine. Not sure? Get 500 GB. Storage is the one thing here you can't work around later.
Same build, same day. No cut-down version and nothing saved for the expensive machines. The only exception is AI features that need a neural engine, because this chip doesn't have one.
The memory is soldered, so no. The 500 GB model uses an M.2 module you could physically swap, but it's not a machine you're meant to service yourself and we won't pretend it is.
Yes, it's the machine we'd put in a classroom. Silent, light, cheap to replace, and running a system that doesn't watch whoever's using it. What that looks like at scale is on the districts page.
Release status: the AeroBook™ Nano hasn't been manufactured. It's designed, specced and priced, and the $499 in the builder is the price we plan to keep. But no unit has been built, nothing is in production, and we're not taking orders or giving a ship date. Treat every number on this page as a design target, not a measurement. Intel and the Intel logo are trademarks of Intel Corporation. Noctic has no partnership with or endorsement from Intel.
A desktop computer fifteen centimetres square. Sixteen cores, 32 GB of memory and four USB-C ports, in a box light enough to carry from one desk to the next.
A concept. Prices are indicative, and buying is locked until 2030.
From $1,499, indicative. Buying opens at release in 2030.
Most of a desktop computer is air. The Aero™ Neo is what's left when you take the air out and keep the processor.
It has the same Intel Core Ultra 9 285H as the AeroBook™ Pro, minus the screen, keyboard and battery, which is most of what a laptop costs. Plug it into the display you already have, grab it at the end of the day and plug it into another one. It runs off a USB-C charger, so the second desk just needs a cable, not a second computer.
Development, office and school work with a ton open, photo editing, 4K video, virtual machines, and anyone who works at more than one desk.
Working on a train. There's no screen and no battery, and a desktop that pretends otherwise is just a laptop with extra cables.
A fan you'll hear when all sixteen cores are busy, graphics that are good for integrated but aren't a graphics card, and 32 GB that's fixed the day it's made.
A 140 W USB-C adapter comes in the box. On a smaller charger the Neo still starts and still works. It just holds back under heavy load, and it tells you.
No display, keyboard or mouse in the box. Anything that speaks USB-C or Bluetooth works, and one cable to a USB-C monitor carries the picture as well.
The Aero™ has eight more cores, twice the memory and a fan that barely turns. It also weighs four kilos and stays where you put it. The Neo is the one that comes with you.
An Arrow Lake H-series chip with sixteen cores and memory soldered right next to it, cooled by one fan tuned to stay slow, not to break records.
| Processor | Intel® Core™ Ultra 9 285H |
|---|---|
| Platform | Core Ultra 200H series |
| Cores / threads | 16 (6 performance + 8 efficiency + 2 low-power) / 16 |
| Max turbo | 5.4 GHz |
| Cache | 24 MB L3 |
| Graphics | Intel® Arc™ 140T, 8 Xe cores |
| Neural engine | Intel® NPU 3, 13 TOPS |
| Memory | 32 GB LPDDR5X-8400, soldered to the board |
| Upgradeable | No. Nothing socketed, nothing removable. |
| Intended successor | N1 Bionic (designed, not yet made) |
No battery to protect and no lap to keep cool, so the Neo can run the processor harder than the AeroBook™ Pro does, and for longer.
Arc 140T shares the 32 GB with the processor. It's quick for photos, video and light 3D, but it's not a graphics card. If your work depends on one, this isn't your machine.
A vapour chamber sits on the processor. It's a sealed plate with a little liquid inside that boils off the hot chip and condenses at the cool edges, moving heat like a tiny fridge with no pump. One blower, tuned slow, pushes that heat out of the case: silent while you write, audible during a long export.
N1 Bionic: eight cores and 16 GB on the package, designed for exactly this kind of small, efficient machine. It doesn't exist. The plan is on Noctic Silicon.
Six centimetres tall on purpose, because sixteen cores need a real heatsink and room for a fan to spin slowly. One piece of aluminium on a dark foot, the Aero™ mark on top and a status light on the front. Two USB-C ports face you, two face the wall, and power goes into whichever's closest.
| Size | 15 × 15 × 6 cm |
|---|---|
| Weight | 1.5 kg |
| Ports | Four USB-C (two front, two rear), USB4 at 40 Gb/s |
| External displays | Up to three, at up to 6K |
| Power | USB-C Power Delivery, 140 W adapter in the box |
| Wireless | Wi-Fi 7 and Bluetooth 6.0 |
| Wired network | None built in. The Aero™ Dock adds Ethernet from one port |
| Storage | 1 TB NVMe, around 7 GB/s sequential read, with 2 TB and 4 TB to order |
| Cooling | One quiet fan over a vapour chamber |
| Audio | One small speaker, for alerts. No microphone. |
| Body | Aluminium unibody, in Silver or Graphite |
USB-C only, like every machine in the family. If the second desk needs HDMI, USB-A, an SD slot or Ethernet, one dock covers all of it and stays put when the Neo leaves.
Planned figures for an unbuilt machine. Every one of them may change.
| Model | Aero™ Neo |
|---|---|
| Processor | Intel® Core™ Ultra 9 285H |
| Cores / threads | 16 (6P + 8E + 2LP-E) / 16 |
| Max turbo | 5.4 GHz |
| Cache | 24 MB L3 |
| Graphics | Intel® Arc™ 140T, 8 Xe cores |
| Neural engine | Intel® NPU 3, 13 TOPS |
| Memory | 32 GB LPDDR5X-8400, soldered |
| Storage | 1 TB NVMe, with 2 TB and 4 TB options |
| Ports | 4 × USB-C, USB4 40 Gb/s |
| External displays | Up to three, up to 6K |
| Wireless | Wi-Fi 7, Bluetooth 6.0 |
| Power | USB-C Power Delivery, 140 W adapter included |
| Cooling | One fan, vapour chamber |
| Size and weight | 15 × 15 × 6 cm · 1.5 kg |
| Body | Aluminium unibody |
| System | AXESS OS: Linux-based, Glass UI, AppImage support |
| Telemetry | None unless you switch it on |
| Warranty | Intended at two years, parts and labour |
| Intended successor silicon | N1 Bionic. See Noctic Silicon |
| Aero™ Neo | Aero™ | AeroBook™ Nano | AeroBook™ Pro | |
|---|---|---|---|---|
| Type | Desktop | Desktop | Laptop | Laptop |
| Processor | Core Ultra 9 285H | Core Ultra 9 285 | N100 | Core Ultra 9 285H |
| Threads | 16 | 24 | 4 | 16 |
| Memory | 32 GB | 64 GB | 8 GB | 32 GB |
| Screen | Bring your own | Bring your own | 12.5″ 60 Hz | 16.2″ 144 Hz |
| Weight | 1.5 kg | 4.1 kg | 1.02 kg | 2.14 kg |
| From | $1,499 | $2,499 | $499 | $3,499 |
| Read more | You're here | Aero™ | AeroBook™ Nano | AeroBook™ Pro |
It weighs 1.5 kg and fits in a bag next to a laptop. It's portable the way a laptop charger is. You carry it between places that already have a screen, a keyboard and a plug.
If you work away from desks, yes. The AeroBook™ Pro costs $2,000 more for the screen, the battery and the titanium, and if you already own a monitor you'd be paying for a second one.
Any USB-C Power Delivery charger will start it. Below 100 W it holds back under heavy load; on the 140 W adapter in the box it doesn't.
No. The 32 GB is soldered beside the processor and there's no larger configuration. If you know you need more, the Aero™ has 64 GB.
No. Every port is USB-C. The Aero™ Dock adds Ethernet, HDMI, three USB-A ports and SD off one cable, and Wi-Fi 7 is fast enough that most people won't miss the socket.
Concept status: the Aero™ Neo is a concept. Nothing's been manufactured, there's no manufacturing partner and no release date. The prices in the builder are indicative and buying is locked until release, which we don't expect before 2030. Every number on this page is planned, not final, and could change or get dropped. Intel, Core Ultra and Arc are trademarks of Intel Corporation. Noctic has no partnership with or endorsement from Intel.
The full-size desktop. Twenty-four cores, 64 GB of memory and six USB-C ports, in a plain aluminium case with one big fan that barely has to turn.
A concept. Prices are indicative, and buying is locked until 2030.
From $2,499, indicative. Buying opens at release in 2030.
A desktop gets to be bigger than it seems like it needs to be, and the Aero™ uses that room on the two things a small machine can't have: more cores, and quiet while they work.
It's built around Intel’s Core Ultra 9 285, a desktop chip with eight performance cores and sixteen efficiency cores, with 64 GB of memory next to it. The case is sized around one 140 mm fan, because a big fan moving slowly is the only way to cool a chip like this without it sounding like one.
Large compiles, virtual machines, rendering on the processor, audio production, big photo libraries, local models that live in memory, and long jobs you'd rather not listen to.
Games and anything that needs a graphics card. The Aero™ has integrated graphics and no slot for a card, and no amount of processor makes up for that.
Four kilos, a 240 W USB-C adapter and a spot on the desk it never leaves. Memory and storage are fixed the day it's made, like every machine in the family.
A vapour chamber pulls the heat off the processor, and one 140 mm fan clears it out. The chamber is a sealed plate with a little liquid inside that boils off the hot chip and condenses at the cool edges, moving heat like a tiny fridge with no pump. Most of the time the fan turns too slowly to hear. Under a long export you'll notice air moving, not a whine.
No display, keyboard or mouse in the box. Up to four displays connect over USB-C, one of them at 8K, and anything Bluetooth pairs from the first screen you see.
The Aero™ Neo is a thousand dollars less, a tenth of the size, and fast enough for most people. Get the Aero™ when you can name the job that needs twenty-four cores.
An Arrow Lake desktop chip: eight performance cores for work that has to finish now, and sixteen efficiency cores for everything that can run next to it.
| Processor | Intel® Core™ Ultra 9 285 |
|---|---|
| Platform | Core Ultra 200S series (desktop) |
| Cores / threads | 24 (8 performance + 16 efficiency) / 24 |
| Max turbo | 5.6 GHz |
| Cache | 36 MB L3 |
| Base power | 65 W |
| Graphics | Intel® Graphics, 4 Xe cores, integrated |
| Neural engine | Intel® AI Boost NPU, 13 TOPS |
| Memory | 64 GB DDR5-5600, soldered to the board |
| Upgradeable | No. Memory is soldered, there are no slots, and the case isn't made to be opened. |
| Intended successor | N1 Pro (specification only) |
The Core Ultra 9 285K adds a tenth of a gigahertz and sixty more watts of base power. In a case built to be quiet, that just buys you noise, not speed.
Unusual for a machine this size, and on purpose. It keeps the family’s rule that nothing inside changes after it's made. Buy the size you'll still want in five years.
Four Xe cores are enough for a few big displays, video and a desktop that never stutters. For 3D and games they're way behind a graphics card, and we'd rather just say that here.
N1 Pro: twenty-four cores with memory on the package, the same eight-and-sixteen split. It doesn't exist and it's the hardest part to build. The plan is on Noctic Silicon.
One aluminium shell with nothing on the front but the Aero™ mark and two USB-C ports. Air comes in through the gap above the foot and goes out the back, where the power button and four more USB-C ports are.
| Size | 16 × 16 × 27 cm |
|---|---|
| Weight | 4.1 kg |
| Ports | Six USB-C (two front, four rear), USB4 at 40 Gb/s |
| External displays | Up to four, one of them at 8K |
| Power | USB-C Power Delivery, 240 W adapter in the box |
| Wireless | Wi-Fi 7 and Bluetooth 6.0 |
| Wired network | None built in. The Aero™ Dock adds Ethernet from one port |
| Storage | 2 TB PCIe 5.0 NVMe, around 14 GB/s sequential read, with 4 TB and 8 TB to order |
| Cooling | One 140 mm fan over a vapour chamber |
| Audio | One speaker, for alerts. No microphone. |
| Body | Aluminium, in Silver or Graphite |
Power comes in over USB-C like everything else. USB Power Delivery carries up to 240 W, which is enough to run the Aero™ flat out. Need HDMI, USB-A, SD or Ethernet? One dock gives you all of it.
Planned figures for an unbuilt machine. Every one of them may change.
| Model | Aero™ |
|---|---|
| Processor | Intel® Core™ Ultra 9 285 |
| Cores / threads | 24 (8P + 16E) / 24 |
| Max turbo | 5.6 GHz |
| Cache | 36 MB L3 |
| Base power | 65 W |
| Graphics | Intel® Graphics, 4 Xe cores |
| Neural engine | Intel® AI Boost NPU, 13 TOPS |
| Memory | 64 GB DDR5-5600, soldered |
| Storage | 2 TB PCIe 5.0 NVMe, with 4 TB and 8 TB options |
| Ports | 6 × USB-C, USB4 40 Gb/s |
| External displays | Up to four, one at 8K |
| Wireless | Wi-Fi 7, Bluetooth 6.0 |
| Power | USB-C Power Delivery, 240 W adapter included |
| Cooling | One 140 mm fan, vapour chamber |
| Size and weight | 16 × 16 × 27 cm · 4.1 kg |
| Body | Aluminium |
| System | AXESS OS: Linux-based, Glass UI, AppImage support |
| Telemetry | None unless you switch it on |
| Warranty | Intended at two years, parts and labour |
| Intended successor silicon | N1 Pro. See Noctic Silicon |
| Aero™ Neo | Aero™ | AeroBook™ Nano | AeroBook™ Pro | |
|---|---|---|---|---|
| Type | Desktop | Desktop | Laptop | Laptop |
| Processor | Core Ultra 9 285H | Core Ultra 9 285 | N100 | Core Ultra 9 285H |
| Threads | 16 | 24 | 4 | 16 |
| Memory | 32 GB | 64 GB | 8 GB | 32 GB |
| Screen | Bring your own | Bring your own | 12.5″ 60 Hz | 16.2″ 144 Hz |
| Weight | 1.5 kg | 4.1 kg | 1.02 kg | 2.14 kg |
| From | $1,499 | $2,499 | $499 | $3,499 |
| Read more | Aero™ Neo | You're here | AeroBook™ Nano | AeroBook™ Pro |
If you're not sure, get the Neo. The Aero™ is for people who already sit waiting on their computer: long compiles, lots of virtual machines, renders that take hours.
No. There's no slot, and the case is sized to cool the processor quietly, not to fit a card. If your work runs on a graphics card, this isn't the machine for it.
Because every machine in the family works that way, and because 64 GB is picked to last the whole life of the machine. It's a real limit, and it's why there's only one size.
Close to silent doing normal stuff at a desk. Under a long full load you'll hear air moving. One big fan is the quietest way we know to cool twenty-four cores.
Yes. USB Power Delivery carries up to 240 W, and the Aero™ stays under that at full load. A smaller charger still starts it. The machine just holds back under heavy work until it has enough.
Concept status: the Aero™ is a concept. Nothing's been manufactured, there's no manufacturing partner and no release date. The prices in the builder are indicative and buying is locked until release, which we don't expect before 2030. Every number on this page is planned, not final, and could change or get dropped. Intel and Core Ultra are trademarks of Intel Corporation. Noctic has no partnership with or endorsement from Intel.
Sixteen cores, a 16.2-inch 144 Hz screen, four USB-C ports at 80 Gb/s, and a vapour chamber built so the chip never has to slow down.
A concept. Prices are indicative, and buying is locked until 2030.
From $3,499, indicative. Buying opens at release in 2030.
If you're reading this page to figure out whether you need a Pro, you probably don't. The people this is for already know exactly what they need, because their work told them.
Sixteen cores (six performance, eight efficiency, two low-power) in a body with the room and airflow to keep all of them busy for an hour without backing off. Heavy software, heavy gaming, large models, big compiles, long exports. It weighs 2.14 kg and you'll hear the fan when you push it. That's the point, not a flaw.
Long renders, big compiles, video above 4K, 3D, heavy gaming, local models, and any job where the machine's busy longer than it's idle.
Twice the weight of the AeroBook™ Nano, $2,000 more than an Aero™ Neo with the same chip, and a fan you'll hear under load. That's exactly what performance you can carry costs.
Six performance cores for work that has to finish now, eight efficiency cores for throughput, and two low-power cores for everything in the background. One thread each, on purpose.
More than twice the Nano’s battery for two more hours, because this machine actually pulls power. That's just math, not a flaw, and it's why the number says 16 and not something bigger.
Four USB4 v2 ports at 80 Gb/s. This is the machine that turns into a desktop when it gets to a desk, and the ports are what make that work.
The neural engine. The H-series chip has NPU 3 at 13 TOPS, well behind the newest low-power chips. On CPU and graphics the Pro is way ahead of them though.
An Arrow Lake H-series chip: more cores, more cache, more sustained power, and memory soldered to the board, not on the package.
| Processor | Intel® Core™ Ultra 9 285H |
|---|---|
| Platform | Core Ultra 200H series |
| Cores / threads | 16 (6 performance + 8 efficiency + 2 low-power) / 16 |
| Max turbo | 5.4 GHz |
| Cache | 24 MB L3 |
| Graphics | Intel® Arc™ 140T, 8 Xe cores |
| Neural engine | Intel® NPU 3, 13 TOPS |
| Memory | 32 GB LPDDR5X-8400, soldered to the board |
| Upgradeable | No. Soldered, with nothing socketed anywhere in the machine. |
| Intended successor | N1 Pro (specification only) |
The H-series doesn't carry memory on the package, so the Pro solders it right next to the chip instead. Same result for you, fixed at purchase, just a slightly different layout inside.
Not even here. A card means a second pool of memory, copying between the two, and a power budget that costs more than it gives back in a machine this size. The graphics share the 32 GB.
A large vapour chamber over the processor feeds two fans and two intakes. The chamber is a sealed plate with a little liquid inside that boils off the hot chip and condenses at the cool edges, moving heat like a tiny fridge with no pump. It's there so a forty-minute export runs as fast in minute forty as in minute one.
N1 Pro: 24 cores, 5.7 GHz, 24 GPU cores, 205 GB/s. It doesn't exist yet and it's the hardest of the three to build. The plan is on Noctic Silicon, including how it could fail.
The biggest laptop in the family, in a titanium-and-aluminium unibody, with six speakers, two subwoofers and a studio mic array. 16.8 mm thick, because the cooling has to fit somewhere.
| Display | 16.2-inch edge-to-edge 8K QLED concept panel, 1–144 Hz adaptive |
|---|---|
| Brightness | 1,900 nits peak |
| Colour | P3 wide gamut, 10-bit, factory calibrated |
| Touch | Full multi-touch, with touch unlock in the keyboard deck |
| Camera | 12 MP in the Dynamic Notch, with face unlock |
| Keyboard | Backlit, 1 mm travel, with the Aero™ key |
| Speakers | Six, with dual subwoofers |
| Microphones | Four, studio array |
| Ports | Four USB-C, USB4 v2 at 80 Gb/s, charge or data on any |
| External displays | Three 6K, or one 8K |
| Wireless | Wi-Fi 7 and Bluetooth 6.0 |
| Storage | 2 TB NVMe, around 7.4 GB/s sequential read |
| Battery | 100 Wh, up to 16 hours |
| Cooling | Vapour chamber and two fans |
| Weight and thickness | 2.14 kg · 16.8 mm |
| Body | Titanium-and-aluminium unibody |
100 Wh is the biggest battery an airline lets you carry, so that's where the number stops, not wherever would look nicer on a spec sheet.
Planned figures for an unbuilt machine. Every one of them may change.
| Model | AeroBook™ Pro |
|---|---|
| Processor | Intel® Core™ Ultra 9 285H |
| Cores / threads | 16 (6P + 8E + 2LP-E) / 16 |
| Max turbo | 5.4 GHz |
| Cache | 24 MB L3 |
| Graphics | Intel® Arc™ 140T, 8 Xe cores |
| Neural engine | Intel® NPU 3, 13 TOPS |
| Memory | 32 GB LPDDR5X-8400, soldered |
| Storage | 2 TB NVMe |
| Display | 16.2″ 8K QLED, 1–144 Hz adaptive, 1,900 nits |
| Ports | 4 × USB-C, USB4 v2 80 Gb/s |
| Wireless | Wi-Fi 7, Bluetooth 6.0 |
| Battery | 100 Wh · up to 16 h |
| Cooling | Vapour chamber and two fans |
| Weight | 2.14 kg |
| Thickness | 16.8 mm |
| Body | Titanium-and-aluminium unibody |
| System | AXESS OS: Linux-based, Glass UI, AppImage support |
| Telemetry | None unless you switch it on |
| Warranty | Intended at two years, parts and labour |
| Intended successor silicon | N1 Pro. See Noctic Silicon |
| Aero™ Neo | Aero™ | AeroBook™ Nano | AeroBook™ Pro | |
|---|---|---|---|---|
| Type | Desktop | Desktop | Laptop | Laptop |
| Processor | Core Ultra 9 285H | Core Ultra 9 285 | N100 | Core Ultra 9 285H |
| Threads | 16 | 24 | 4 | 16 |
| Memory | 32 GB | 64 GB | 8 GB | 32 GB |
| Screen | Bring your own | Bring your own | 12.5″ 60 Hz | 16.2″ 144 Hz |
| Weight | 1.5 kg | 4.1 kg | 1.02 kg | 2.14 kg |
| From | $1,499 | $2,499 | $499 | $3,499 |
| Read more | Aero™ Neo | Aero™ | AeroBook™ Nano | You're here |
If you're not sure, no. If you work at a desk, the Aero™ Neo has the same chip for less than half the price. The Pro is for people whose heavy work has to travel.
Because it has sixteen cores and cooling that actually lets them run. A big battery in a machine that pulls real power empties faster than a small one in a machine that can't.
No. Integrated Arc 140T sharing 32 GB. For rendering and compute that's a good trade. For top-end gaming it isn't, and if that's your thing, this isn't your machine.
You'll hear it under long heavy load, and it's silent the rest of the time. A machine this fast either makes noise or slows down, and we'd rather it made noise.
Not this generation. The memory is soldered at 32 GB and there's no bigger option. If your work needs 64 GB and can stay on a desk, the Aero™ has it.
Concept status: the AeroBook™ Pro is a concept. Nothing's been manufactured, there's no manufacturing partner and no release date. The prices in the builder are indicative and buying is locked until release, which we don't expect before 2030. Every number on this page is planned, not final, and could change or get dropped. Intel, Core Ultra and Arc are trademarks of Intel Corporation. Noctic has no partnership with or endorsement from Intel.
Four machines, and the right answer usually isn't the expensive one. Answer seven questions and we'll tell you which one we'd point you at, why, and where we'd argue with ourselves.
The quiz is there to help, it's not magic. If you already see yourself in one of these lines, that's your machine.
| Aero™ Neo | You work at a desk, or a few of them, and want real speed in something small and quiet. You've already got a screen, or don't mind picking one. |
|---|---|
| Aero™ | Your work already decided for you: long compiles, virtual machines, renders on the processor, all at one desk. You want twenty-four cores and 64 GB, and you don't need a graphics card. |
| AeroBook™ Nano | Price is what decides it. It browses, writes, does calls and goes in a bag. You're fine with 8 GB, four threads and a 60 Hz screen because it costs so little. |
| AeroBook™ Pro | Heavy work that has to travel: long renders, big compiles, local models away from a desk. And you're fine carrying 2.14 kg and hearing a fan for it. |
| None of them | You need a dedicated GPU, more than 64 GB of memory, or a machine you can upgrade later. All three are limits we chose for this family, and no configuration gets rid of them. |
Pick the machine, the finish, how much storage you want and everything that goes with it. Prices are indicative. Nothing can be bought until the machines exist, which we don't expect before 2030. Whatever you build here can be reserved.
Aero™ Neo
Aluminium Silver · 15 cm square · 6 cm tall · 1.5 kg
A configurator with a locked button is a weird thing to publish. It's here because the machines are designed down to the storage options, and we'd rather show you the real shape of the product than put up a page that pretends to sell something.
Every option here is a real design decision. What you pick tells us which ones matter.
Nothing's quoted, tooled or on a production line yet. Until that changes, we can't commit to any price here.
Once there's a bill of materials, these numbers get replaced with ones we can actually stand behind.
We ask reservations to confirm first. If you say nothing, nothing happens.
Because a spec with no price tells you nothing about whether the machine's for you. These are the numbers we're designing toward, marked indicative everywhere they show up, and they'll move.
No. No card, no deposit, no address. It saves which machine you're into and holds you a spot in the first hundred. The reservation page explains the whole thing.
The build you send with a reservation is saved to your Noctic Account so we know what to make first. It's not a commitment for either of us, and you can change it whenever.
Because that's a realistic read of how long a machine like this takes with no manufacturing partner yet. If it moves, it'll move later, not earlier, and we'll say so on the changelog.
Our own RISC-V processor, designed here, for our own machines and our own operating system. It's the hardest thing we've ever proposed. It's not impossible, and this page is the whole plan, including the parts most likely to kill it.
Designed: N1 Bionic. See the chip
Noctic designs its own semiconductors. We want to build a processor. Not a chip someone else designed with our name on the lid. A processor we specify, architect, verify and get manufactured. It uses the RISC-V instruction set, which is open: no architecture licence to sign, no royalty to pay, and room to extend it where the system needs something the base doesn't have. The cost is software. Anything that can be built from source runs once it's compiled for RISC-V, and a binary that only ever shipped for x86 doesn't run at all. We have designed it: the first processor, N1 Bionic, is specified end to end, and its core is written in SystemVerilog and runs compiled C in RTL simulation. We're a fabless chip designer, the same way Apple, AMD and Nvidia are: we design the chip, and a foundry makes it. What it doesn't have yet is silicon, and the two things that decide when it gets some are patents and money.
Everything below is written as though it'll happen. Read the list of ways it dies before you believe any of it.
Designed. Running. Not yet made.
WE HAVE IT DESIGNED. The architecture is specced end to end and the core runs real programs in simulation. A small test-chip version is being laid out on the open SkyWater 130 nm process. Nothing's been manufactured yet.
Design targets. None of these has been measured.
Designed, but still to be built as hardware: early mispredict recovery, out-of-order memory, real caches, the full width, floating point and vectors, then the MMU, the efficiency cores, the shared caches, the on-chip network, the neural engine and the security part. The performance targets aren't proven until a cycle-level model runs real workloads.
The planned floorplan, drawn in text. Move your pointer over it to light up a part of the chip.
Block by block, with where each one stands. In RTL means it’s written in SystemVerilog and runs in simulation. Designed means it’s specified but not written as hardware yet. Target means it’s a number we’re aiming for. Nothing here has been measured on silicon, because there isn’t any yet.
| Block | How it works | Where it stands |
|---|---|---|
| Instruction set | RV64IM today: the 64-bit integer base plus multiply and divide. The target is RV64GCBV on the RVA23 profiles, adding compressed instructions, atomics, single and double precision floating point, bit manipulation and RVV 1.0 vectors with 256-bit registers. | In RTL: RV64IM. The rest is designed. |
| Branch direction | TAGE. A base table indexed by the branch address, plus tagged tables indexed by a hash of the address and the global branch history, each table using a longer history than the last. The longest matching history makes the prediction. A miss allocates into a longer-history table, and usefulness counters decide which entries can be replaced. | In RTL. Checked against an independent model under constrained random stimulus. |
| Rename | Eight instructions a cycle. A register alias table maps each architectural register to the physical register holding its newest value, and a free list hands out unused ones. Inside a group of eight, each instruction checks every older one for a write to its sources: 28 comparisons per source operand, 56 for two, all in one cycle. That cost grows with the square of the width. | In RTL, model-checked the same way. |
| Reorder buffer | 512 entries in program order. Instructions enter at rename and leave at retire, strictly in order, so exceptions are precise. Today a mispredicted branch is recovered when it reaches the head of the buffer, by restoring the rename table from the retired state. Early recovery, checkpointing the table at every branch, is designed. | Buffer in RTL, model-checked. Early recovery designed. |
| Issue | Instructions wait until their operands are ready. Every cycle, finishing operations broadcast their destination tag (wakeup), each waiting instruction compares it with its sources, and ready instructions are picked for the execution ports (select). That wakeup-and-select loop is one of the hardest timing paths in any out-of-order core. | In RTL, model-checked. |
| Loads and stores | A store queue and a load queue. A load searches older stores for the same address and takes the data straight from the queue when it finds it, which is store-to-load forwarding. Running loads ahead of stores whose address isn’t known yet, with a memory dependence predictor and replay, is designed. Ordering follows RISC-V’s weak memory model, RVWMO. | Forwarding in RTL. Out-of-order memory designed. |
| Caches | 32 KB 8-way L1 instruction and data caches, virtually indexed and physically tagged. 256 KB private L2. A 16 MB L3 sliced around a ring, 64-byte lines, MESI coherence, miss status registers so each cache keeps several misses in flight, and stride and stream prefetchers. | Designed. Target sizes. |
| Address translation | Sv39 and Sv48 page tables, small TLBs beside each L1 with a bigger shared TLB behind them, and a hardware page-table walker. | Designed. |
| Floating point and vectors | IEEE 754 single and double precision, and the RVV 1.0 vector unit. | Designed. |
| Efficiency cores, fabric, neural engine, security | Four efficiency cores sharing an L2, the on-chip network, a 12.3 TOPS neural engine and the root of trust. Graphics would be licensed IP. | Designed. Targets. |
Cycles per instruction of 0.679 on Dhrystone means 1/0.679, about 1.47 instructions finished every cycle on average. An in-order core can’t pass 1, because every stall stops everything behind it. Going past 1 is the out-of-order machinery doing its job.
At eight a cycle, a 512-entry window fills in 512 / 8 = 64 cycles. That covers an L2 or L3 hit. It doesn’t cover a trip to DRAM, which takes a few hundred cycles, but it does let the core find the next few misses while the first one is waiting. Overlapped misses cost about the same as one, and that overlap is called memory-level parallelism.
32 KB / 8 ways = 4 KB per way, which is exactly one 4 KB page. So every index bit sits inside the 12-bit page offset that translation never changes. The cache can start its lookup before the TLB answers, and the same line can never end up stored under two names.
LPDDR5X-8533 moves 8,533 million transfers a second. On a 128-bit bus that’s 8,533 × 128 / 8 = 136.5 GB/s. N1x is LPDDR5X-6400 on 128 bits: 102.4 GB/s. N1 Pro’s 205 GB/s needs 6400 on 256 bits or 8533 on 192, and that bus isn’t chosen yet.
TOPS = 2 operations per multiply-accumulate × the number of MAC units × the clock. For scale, 4,096 INT8 MACs at 1.5 GHz is 12.3 TOPS. The array size and clock behind N1 Bionic’s target haven’t been published, so read it as a target, not a design.
Dynamic power goes with capacitance × voltage² × frequency, and a higher clock needs a higher voltage. So the last few hundred megahertz cost far more power than they add speed. That’s why 4.2 GHz sits inside a 30 W burst budget and the chip lives at 15 W sustained.
Dhrystone is small and old, it fits in any cache, and compilers know it well. Running it correctly at 2.162 DMIPS/MHz proves the pipeline, the predictor and the forwarding work together on real compiled C. It doesn’t predict how fast a laptop would be. CoreMark and SPEC CPU 2017 on a cycle-level model with real caches would, and we haven’t run those yet. When we do, the results go here, good or bad.
The test-chip configuration is being laid out on SkyWater’s SKY130, a 130 nm process with an open process design kit. Open PDK, open-source tools, so the whole flow from RTL to GDSII runs without a licence: synthesis, floorplan, place and route, clock tree, static timing, DRC and LVS. It’s slow and old on purpose. The point is proving the flow end to end, not speed. Nothing has been taped out.
A factory has to tell a good die from a bad one in seconds, so design-for-test goes in from the start. Scan chains link the flip-flops into shift registers a tester can load and read. Built-in self-test covers every memory array. A debug port follows the RISC-V debug specification, so on real silicon we can halt the core, read its registers and step it one instruction at a time.
An instruction set is the deal between software and silicon. Picking one decides, on day one, which programs already run and which have to be rebuilt, translated or dropped. x86 is covered in patents at the modern end and its decoder alone is a research project. Arm means paying Arm and designing inside their licence. We're taking the third option, and with it the software problem, not the licensing one.
| x86-64 | Arm | RISC-V | |
|---|---|---|---|
| Software you already own | Runs. Every Linux binary, every AppImage, every proprietary tool, unchanged. | Most of it needs rebuilding, and the closed-source stuff needs the vendor to care. | Some of it. The desktop side is thin and filling in slowly. |
| Licensing | No open licence. Patent-encumbered at the modern end. The hardest of the three. | Available, but you pay Arm and you build inside their architecture licence. | Free and open. Nothing to sign, nothing to pay. |
| Firmware and boot | One well-trodden path: UEFI or coreboot, ACPI, standard PC platform. | All over the place. Every device tree is its own little archaeology project. | Getting better fast, still not mature for everyday laptops. |
| Virtualising Windows | Native speed, no translation layer. | Possible with Windows on Arm, with caveats. | Not realistically. |
| Difficulty of building one | Highest. The decoder alone is a research project. | High. | Lowest. You can build a working core as a student. |
| Our verdict | RISC-V. It's the only one of the three a company our size can build and own outright. Nothing to licence, nothing to pay, and room to add what AXESS OS needs. The cost is that thin column on the right, and closing that gap is our problem to solve, not a reason to rent someone else's architecture. | ||
AXESS OS is Linux, and Linux runs on RISC-V. What's thin is the long tail, the one tool somebody needs that nobody's rebuilt yet. Checking for that is exactly the homework Aero™ is supposed to take away, so carrying the distribution is our job, not yours.
Emulating one architecture on another works. Rosetta and FEX both proved it. But it costs performance, battery, and a whole class of bugs that only show up on your machine. Native code running natively has none of that.
When the same company owns the scheduler and the core, the two can be built to agree. Which work goes on an efficiency core, what the power states mean, how the neural engine gets fed: today those are guesses. If we build both, they're decisions.
Putting memory inside the processor package and not in slots is what makes the bandwidth and latency numbers on the Aero™ page possible. You can't get that buying a chip off a price list. It's a packaging decision, and those belong to whoever designs the part.
This is the most useful thing to get. Inside, a modern x86 processor is already a fast, simple, RISC-style machine. Bolted on the front is a decoder that reads x86 instructions and translates them, in hardware, into the simple operations the machine actually runs. That decoder never goes away. It's silicon, power and complexity spent staying compatible with a format from 1978. Building on RISC-V is the same machine without the translator on the front.
Pull bytes of program from the instruction cache, guided by the branch predictor guessing where the code goes next.
Read 16- and 32-bit instructions straight out of the stream. The low two bits give the length, so there’s no hunting for where one ends and the next begins. Each one becomes one or more micro-operations.
Map the 32 integer and 32 floating-point registers the program can see onto a much bigger pool of physical registers, so instructions that only reuse a register name stop waiting on each other.
Hold micro-ops until their inputs are ready, then issue them in whatever order the machine can actually run them.
Integer, vector, branch and address units do the work, several at once, across a set of parallel ports.
Loads and stores line up, check each other for conflicts, and hit L1, L2, L3 or main memory, in that order of hope.
Results are committed strictly in program order, so from the outside all that chaos looks exactly like a machine doing one thing at a time.
The x86 decoder is where the difficulty lives. Instructions are between one and fifteen bytes long, and you can't know where the second one starts until you've finished parsing the first. Arm's 64-bit instructions are always four bytes. RISC-V's are four bytes, or two with the compressed extension, and the lowest two bits of every instruction say which, so finding where each one starts takes a few gates instead of a search. Solving that, usually with parallel speculative decoders and a cache of already-decoded micro-ops, is most of what makes an x86 front end expensive.
One level down
The floorplan above shows which blocks exist. This one shows how they reach memory, which (as the cache note below explains) is where almost all real-world speed is won or lost.
| Cores and threads | 8 / 8, no SMT |
|---|---|
| L1 per core | 32 KB instruction + 32 KB data |
| L2 per core | 256 KB, private |
| L3, shared | 16 MB, 2 MB per core |
| Interconnect | Ring bus, 64-byte line size |
| Memory interface | On-package LPDDR5X, 128-bit: eight 16-bit channels |
| Die size target | ~120–150 mm² |
| Package power target | Set per part. See the parts |
| Status | Target figures for a planned part. Nothing here has been taped out, verified on silicon or manufactured. |
A performance core is about four times the size of an efficiency core for way less than four times the work. If something isn't urgent, running it on the small core finishes it at the same moment for a fraction of the energy. The scheduler decides which is which, and that scheduler is ours.
Simultaneous multithreading fills empty slots by pretending one core is two. It wins on servers. On a laptop it buys throughput and sells predictability, and it's been behind ten years of side-channel bugs. We'd rather be consistent than impressive.
A miss all the way to memory costs a few hundred cycles. Almost all real-world speed is won by not missing: bigger caches, better prefetchers, smarter replacement. Clock speed is the number on the box. Cache is the number you feel.
A separate card means a second pool of memory and copying back and forth all the time. One shared pool is slower at the very top end and better almost everywhere else, and it uses way less power. You only fully get that choice if you design the package.
The chip guesses which way branches go and runs ahead on the guess. When it's wrong it throws the work away. Spectre and Meltdown showed that the thrown-away work leaves fingerprints. Any new core has to be designed against that from the first sketch, not patched after.
x86 chips patch themselves with microcode. A RISC-V core like this one is mostly hardwired and has no big microcode store, so field fixes work two other ways. Chicken bits are control registers that switch off a feature that misbehaves, like a prefetcher or a fast path. Machine-mode firmware can trap an instruction the hardware gets wrong and emulate it in software. Both are easier to verify than microcode, and we’d document both publicly.
Chip design isn't one job. It's about a dozen different jobs that hand work to each other in a fixed order. A mistake caught at step four costs an afternoon. The same mistake caught at step nine costs a year and a mask set.
What the part has to do, at what power, how hot, at what price, by when. Written down and argued over before anyone draws a single block.
A software simulator of the microarchitecture, cycle-approximate, running real workload traces. This is where cache sizes, pipeline depth and issue width get decided, in software, where changing your mind is free.
The actual design: every buffer depth, every port, every queue, every corner case in the decoder. Hundreds of pages of spec that the RTL then implements exactly.
The design written as code, in SystemVerilog, describing what every register holds on every clock edge. It reads like software and behaves like wiring.
Half to two thirds of the whole effort. Constrained-random tests, formal proofs on the tricky blocks, and running the design against real RISC-V programs to check it matches a reference model instruction by instruction.
Tools turn the RTL into real logic gates from the foundry's standard-cell library, and then the long fight over area, timing and power starts.
Millions of standard cells get placed, a clock tree gets built to reach every flip-flop at nearly the same moment, and every net gets routed across a dozen or more metal layers. Timing closure is the part that eats months: every path has to meet setup and hold at every corner of process, voltage and temperature.
Static timing, power integrity, electromigration, design-rule and layout-versus-schematic checks. Scan chains and built-in self-test are added so that a factory can tell a good die from a bad one in seconds.
The design is frozen and sent to the foundry. Photomasks get cut. This is where the project stops being reversible and starts being expensive.
Twelve to sixteen weeks of lithography, etch, implant and deposition, from several hundred process steps on a mature node to well over a thousand at the leading edge, ending in a wafer of dies that are tested while still attached to it.
Good dies are cut out and mounted on a substrate with the memory stacks beside them, connected, lidded and tested again. Advanced packaging is its own supply chain with its own queue.
First silicon arrives and someone tries to make it print a character. Then boot firmware. Then a kernel. Bugs found here turn into an errata entry, a chicken bit or firmware workaround, or a stepping, which is a partial re-spin that costs months and new masks.
Compiler support, kernel support, power management, graphics driver, firmware. AXESS OS has to know the part exists and how to run well on it, and none of that happens by itself.
Thousands of hours across voltage, temperature and ageing, until failure rates are known, not hoped for. Only then does a part go into something a person buys.
None of this is secret. The tools are Cadence, Synopsys and Siemens. The foundries are TSMC, Samsung, GlobalFoundries and Intel. Every fabless company follows the same process. The barrier was never knowledge. It's money, people and time, in that order.
We're not going to try a laptop chip as our first project, because we already know how that ends. The plan is a ladder. Each rung is affordable from the one below it, makes something real, and can be dropped without wasting the rungs underneath.
An RV64GC functional model that defines what correct means, and a cycle-approximate model of the microarchitecture. The functional model is what the hardware gets checked against. The cycle-level model with real caches, running CoreMark and SPEC-style workloads, isn’t finished, and until it is the performance targets stay unproven.
Cost: timeThe N1 Bionic core, written in SystemVerilog. That part exists: it passes the RV64I and RV64M compliance suite and runs Dhrystone in RTL simulation. What’s left on this rung is putting it on an FPGA development board and booting Linux on it, at tens of megahertz. Slow, and proof the core works on real hardware, not just in a simulator.
Cost: low thousands, one to two yearsReal silicon on a mature node, sharing a wafer with dozens of other projects through a multi-project wafer service. Not a product. A few square millimetres that prove we can close timing, hand over a manufacturable design and get working parts back. The first time the design meets physics instead of a simulator.
Cost: tens of thousands to low hundreds of thousandsA modest multi-core part on a mature node, with licensed graphics and I/O blocks instead of our own, running AXESS OS on a development board. Slow by 2026 standards and still useful as a product for something, like a thin client, a home server or a developer box. The first chip anyone outside could hold.
Cost: single-digit millions, a team of dozensOut-of-order cores, hybrid clusters, our own or licensed graphics, memory on the package, an advanced node, real packaging. This is the first rung where we get to say the word competitive, and the first that needs outside money, not revenue.
Cost: tens of millions per tapeout, hundreds of engineersThe chip described further down this page, in an Aero™ Neo, in someone's bag. Leading-edge node, full neural engine, the whole system designed together. Every rung below has to go right for this one to exist at all.
Cost: a different company than we're todayThe realistic read: rungs 0 and 1 are a determined small team with time. Rung 2 is a grant, a partner or a good year. Rung 3 needs a funded business. Rungs 4 and 5 need an industry to decide we're worth backing. We can start climbing without knowing if we reach the top, and every rung is worth having on its own.
Rough public industry numbers, rounded, so the ladder above reads like math and not hope. The mask set is the fixed cost of asking a foundry to make your design at all, before you have a single part you can sell.
| Process node | Mask set, roughly | What it suits | Where it lands for us |
|---|---|---|---|
| 130 nm – 65 nm | Tens of thousands, shared on a shuttle | Test chips, teaching, small controllers | Rung 2 |
| 28 nm | Around one to two million | Low-power SoCs, embedded parts | Rung 3 |
| 7 nm | Around ten to twenty million | Mainstream mobile and laptop silicon | Rung 4 |
| 3 nm and below | Twenty million and up, per tapeout | Flagship processors | Rung 5 |
| And that's only masks | Tool licences, IP blocks, emulation hardware, packaging, qualification and salaries are way bigger. A leading-edge laptop chip program costs several hundred million pounds before the first unit ships. | ||
Three chips, one architecture, so one build of AXESS OS runs on all of them. Think base, Pro and Pro Max: same design, each one a step up from the one before. N1 Bionic is designed and running in simulation. The other two are targets in a spec, not something measured off a chip, and none of the three has been made yet.
| N1x | N1 Bionic | N1 Pro | |
|---|---|---|---|
| Intended for | A fanless machine (none in the lineup yet) | Aero™ Neo | Aero™ and AeroBook™ Pro |
| CPU cores | 4 (2P + 2E) | 8 (4P + 4E) | 24 (8P + 16E) |
| Threads | 4 | 8 | 24 |
| Max turbo | 3.4 GHz | 4.2 GHz | 5.7 GHz |
| L3 cache | 12 MB | 16 MB | 36 MB |
| GPU cores | 4 | 10 | 24 |
| Neural engine | 8-core · 11 TOPS | 16-core · 12.3 TOPS | 32-core · 48 TOPS |
| Memory | 8 GB LPDDR5X, on package | 16 GB LPDDR5X, on package | 16 GB, on package |
| Memory bandwidth | 102 GB/s | 136.5 GB/s | 205 GB/s |
| Instruction set | RV64GC, with the vector extension. | ||
| Multithreading | None. One thread per core, deliberately. | ||
| Package | Soldered, with memory on the substrate. Nothing socketed, nothing upgradeable. | ||
| Status | Specification only. Not designed, not manufactured, no date. | Designed and running in simulation. Not manufactured, no date. | Specification only. Not designed, not manufactured, no date. |
Until these exist, Aero™ runs Intel Core Ultra silicon. It's a real, shipping platform, so the machine and the operating system can be judged now instead of in a decade. What Aero™ uses today.
Aero™ ships on someone else's processor, and it's not because we're impatient. A custom chip is only worth anything if the software for it already exists and is already good. Building the OS on hardware anyone can buy means AXESS OS gets judged on its own, and the day Noctic Silicon works there's something ready to put on it.
The Aero™ machines run Intel today and Noctic Silicon is RISC-V, so this is a real architecture change, not a quiet swap. AXESS OS is built from source, so the system and its packages move over by being rebuilt. What doesn't move over is a closed program that only ever shipped for x86. We'd rather say that up front than have it blow up at the handover.
A custom part means the boot firmware is ours too: coreboot-derived, auditable, and without a management engine nobody outside the vendor has ever seen the source of. That's one of the better arguments for doing any of this.
Hybrid cores only work if the OS knows which thread should get which core. When both sides are ours, that's a decision we make, not a guess based on hardware hints read through a keyhole.
GCC and LLVM already make code for this instruction set. We add a tuning target, not a whole back end. Compare that to a brand new architecture, where just the compiler work takes years before anything runs well.
Every processor ships with bugs. Ours will too. What we can actually do differently is a public errata list in plain language, with what each bug affects and what the fix costs you in performance. Not a PDF nobody links to.
AVX and SSE are x86 instructions, so a RISC-V chip doesn’t have them at all. Code built from source gets the RISC-V Vector extension instead, RVV 1.0, which is vector-length agnostic: the same binary uses whatever width the hardware has, 256 bits on N1 Bionic. Plain loops get auto-vectorised by GCC and LLVM. Hand-written AVX intrinsics don’t, and need porting or a translation header like SIMDe. We’d list exactly which packages that hits.
Written now, while it costs nothing to say, instead of later when it would cost something to admit.
Every rung above rung 2 needs money we don't have and haven't raised. This is the most likely ending: the ladder stops wherever the money stops.
A freedom-to-operate review, where a patent firm reads the design against live claims, could come back saying it can't ship without a licence. And that licence could be unavailable at any price we could pay. That ends it, no matter how good the chip is.
A verification hole that reaches silicon costs a re-spin: three to six months and a fresh mask set. A small company gets one of those. Maybe two.
A first chip four times slower than a shipping Intel part is a cool engineering result and a product nobody buys. There's no partial credit in a laptop.
Leading-edge capacity goes to customers who buy it by the wafer-year. A first-time customer with a small order is at the back of a really long line, if they get in line at all.
There aren't many engineers on earth who've built an out-of-order core of any kind, and almost all of them work for two companies that feel strongly about that. Hiring is harder than funding.
Four to six years per chip means the spec is written for a world that'll have moved on by the time it ships. Aim at where the industry is today and you're guaranteed to show up late.
Intel and AMD make really good x86 chips and sell them to anyone. If the answer at rung 3 is that we can't beat what we can already buy, the right call is to stop, say so on this page, and keep building the software.
A semiconductor design company, yes. Noctic designs its own semiconductors, and N1 Bionic is our own processor design, written in RTL and running in simulation. We’re fabless, like Apple, AMD and Nvidia: we design the chip and a foundry like TSMC or SkyWater makes it. We’ll never be a semiconductor manufacturer with our own factory. When the SkyWater test chip comes back, that’s our first semiconductor chip, and this answer gets updated.
No silicon exists yet. What does exist is N1 Bionic, designed and running real programs in simulation, and the plan on this page. Everything described as a part is a spec, not a measurement.
Because it explains why Aero™ is built the way it is, and because a plan written down in public is one people can hold us to. If the ladder stalls, this page gets updated to say so.
Because the licence problem is permanent and the software problem isn't. Nothing we can do would get us an x86 licence, and an Arm licence would mean designing inside someone else’s rules forever. The distribution we ship is built from source, so most of the gap is a recompile instead of a rewrite.
Not yet, and we're not pretending it's close. The cores exist, Linux runs and the toolchains are good. The gap is desktop polish and the long tail of closed software, and closing that gap is the work.
Intel Core Ultra processors, and an N100 on the Nano. Real parts, available now, with memory soldered right next to the processor. The whole Aero™ design depends on that.
If it's open source, almost definitely. It gets built for RISC-V the same way the rest of the system does. If it's a closed program that only ever shipped for x86, no. That's the real cost of this choice and we're not going to hide it.
There’s no date, and making one up would be lying. Rungs 0 and 1 are under way: the core runs in RTL simulation and its test-chip version is being laid out. A shuttle tape-out, rung 2, needs money we haven’t raised. Rung 5 depends on whether this turns into a very different company.
Concept status: Noctic Silicon is a research and design project. N1 Bionic is designed and runs in simulation, but no processor has been manufactured, verified on silicon or benchmarked on real hardware. The N1x, N1 Bionic and N1 Pro, their core counts, clocks, cache, graphics, neural engine and bandwidth numbers, the process nodes, the costs and the stages on this page are targets and estimates, not commitments, products or final specs. There's no licence agreement, no foundry agreement, no funding and no date.
The assistant that thinks, builds and talks.
Parable 1 reasoning, an agent mode that handles multi-step work, and six effort tiers so you decide how hard it thinks.
Agent mode plans first, then does it. Every step it takes is written down where you can read it.
Give it a goal, not a question. Shiba plans, does the work and reports back.
Economic, Low, Medium, High, Plus and Ultra. Pay only for the thinking you need.
Full projects, multi-file artifacts, live previews and one-click deploys.
Generation, editing, 360° panoramas and style transfer.
Super realistic speech with quick, to-the-point answers.
Remembers what matters to you, and forgets what you tell it to.
Watch it think before it answers.
On effort tiers above Medium, Shiba shows its chain of thought, so you can see why it picked an answer and not only what it picked.
Most assistants decide how hard to think for you. Shiba gives you the control, and shows what each setting costs before you use it.
| Economic | Fast, cheap, literal. Good for lookups, reformatting and anything you already know the shape of. |
|---|---|
| Low | Normal conversation. Enough thinking to be right without making you wait. |
| Medium | The default. Handles most real work, and it's where the reasoning starts showing. |
| High | Multi-step problems, longer code, arguments with more than one moving part. |
| Plus | Work you'd normally block out a whole afternoon for. |
| Ultra | The hardest tier. Slower and pricier on purpose, for problems where being right matters more than being fast. |
Drafting, editing and arguing with your own draft. It'll tell you a paragraph isn't working, not politely rewrite it into something blander.
Whole projects, not snippets: multi-file artifacts, live previews and a deploy at the end. Agent mode runs the build, check, fix loop on its own.
Explaining a concept until it clicks, making practice questions, checking your reasoning. It won't write a student's assignment and pretend that's what was asked.
Deep research goes across lots of sources, keeps track of where each claim came from, and tells you when sources disagree instead of averaging them.
Generation and editing, 360° panoramas and style transfer, in the same conversation as everything else.
Call mode answers at talking speed and gets to the point, because nobody wants an essay read out loud.
Everything Shiba remembers about you is listed in plain sentences, not guessed from some hidden profile.
One thing or all of it, whenever you want. Deleted memory doesn't come back from a cached copy.
Memory is a setting, not a requirement. Conversations work exactly the same without it.
Your conversations aren't used to train models. Memory is for you, not the next model version.
An assistant that'll do anything isn't a feature. What it does depends on the verified age on the account, and the tiers are explained on Safety.
When it refuses, it says why. The full rules are in the Acceptable Use Policy.
Yes, and it's the same account that carries your plan into Cadence. Noctic Account.
Above Medium effort the working shows up next to the answer, so you can check how it got there instead of just trusting it. You can collapse it if you don't want to see it.
Yes, for research and current information, with sources attached to what it tells you.
Your conversations are yours, they're not used to train models, and encrypted mode adds local processing with screenshot and clipboard protection. See Privacy.
Yes. Keys, OAuth and examples are under Developers. Model access goes through our server, so no key ever reaches a browser.
Your campus. Your schedule. Your people.
A social network, calendar, notebook and messenger made for school. In testing on iPhone and Android right now, and anyone at a school can join.
Taken this week from the builds themselves, with the demo timetable in them. The two aren't the same app underneath (iPhone is native, Android runs the web build), so each tab shows its own screens instead of one set pretending to be both.
Cadence AI, powered by Shiba AI, reads the photo, figures out the grid and makes real calendar events you can edit.
Posts, reposts and short video that stay inside your school and grade.
Import a timetable from a PDF or photo. Holidays and PED days included.
AI summaries, cue cards, slideshows and auto-tags. That's Cadence AI, powered by Shiba AI.
School and grade channels, with verified announcements.
Encrypted DMs with local nicknames.
Free blue check by scanning your student ID, verified by AI. The scan is deleted the moment the check finishes. Only the result is kept.
Today's timetable, what changed, and whether that PED day everyone was arguing about is actually happening.
The feed for your grade, the bulletin board for tryouts, and channels that are actually your school.
Notes that turn themselves into cue cards after, tagged to the right subject automatically.
Messages, what's due, and what's coming up tomorrow.
Every school already has one. It works right up until the day it matters.
| You know who is in it | Verification ties an account to a real enrolled student. No former students, no younger siblings, no stranger who got the link in 2023. The card that proves it's deleted as soon as it's read. The feed knows you're a student, and we keep no image of you. |
|---|---|
| Announcements read as official | Only staff can broadcast, channel by channel. A notice from the office actually looks like one. |
| Nothing important scrolls away | The calendar and bulletin board are places you go to, not messages buried under 400 replies. |
| Reports go somewhere | Moderation is done by named people at your school, with a log of every decision and why. |
| Leaving is clean | Switch schools and your access moves with you. Your notes and messages stay yours. |
| Nobody is farming attention | No infinite feed, no streaks, no follower count. Quiet hours are on by default for students. |
It's the first thing every school asks, so here's the answer without the marketing.
Channel posts, bulletin entries, reports filed, and moderation history. Anything published to a group.
Direct messages, private notes and assistant chats. There's no setting that shows them.
Private student content isn't used to train models and isn't read for ads, because there aren't any.
A specific message that's been reported, seen by a moderator only as part of that report.
You can make an account and use notes, messages and your own calendar. Channels and verified announcements need a staff member to claim the campus first.
Take a photo of the printed sheet you got. Cadence AI figures out the grid and makes real calendar events you can edit, not a picture you have to squint at.
That the account belongs to a verified student or member of staff at that school. It's free, and it isn't for sale.
Your calendar and notes stay readable offline, and sync when you're back online.
It installs from the web on any device, with an Android build available. The details are on The base of Cadence. Until release, iPhone and Android test builds go to testers first. Join Cadence Test.
Campus access ends when the school takes you off the roster. The account, your notes and your messages are yours and stay with you.
Try the iPhone and Android apps before they come out.
Until Cadence is released, new builds of the phone apps go to testers first. Sign up here, say which phone you use, and the Cadence team adds you to the test, through TestFlight on iPhone or Google Play’s internal testing on Android. The builds are unfinished on purpose. Finding what breaks is the job.
Which phone, and the email your app store knows you by. That's all we need to add you.
Checking your account…
Cadence runs on a Noctic Account, so testers need one too. It also means only you can change your details or leave the test. It's free and takes a minute.
Sign in or create an accountHere, once. Your place is held against your Noctic Account.
By the Cadence team, through TestFlight or Google Play. It's done by hand, so you'll get an invite later, not a download here.
On iPhone the invite comes from TestFlight. On Android you get a link to join the test on Google Play. Then use it, and tell us what breaks.
The one your phone’s store uses. On iPhone that's your Apple ID (check it in Settings, under your name at the top). On Android it's the Google account signed in to the Play Store. Your school email only works if it's also one of those.
Yes, from 13, or younger with a parent or guardian’s permission. How Cadence treats students and minors is set out in Students and minors, and a test build follows the same rules as the app.
Until Cadence comes out. No date here, because a date we can't keep is worse than no date.
Sign in and press Leave the test on this page. Your sign-up is deleted, the Cadence team is told to take you out of TestFlight or Google Play, and no new builds reach you.
What changed in each build of the phone apps, including what was broken.
Cadence is in testing, so builds go out often and some of them fix what an earlier one broke. Testers get them through TestFlight on iPhone and Play internal testing on Android. Join the test if you want them as they land. Build numbers are the ones shown in Settings, under Help & about.
The wrap camera shows exactly what it records: the whole 9:16 picture, nothing cropped to fill the screen, with 0.5×, 1× and 2× on the back camera, pinch to zoom and tap to focus. Record in clips: stop, start again where the song left off, take back the last clip, then trim or delete any clip before posting. Lyrics can be on your wrap or just for you to lip-sync to.
Add sound shows what’s trending and the sounds people at your school use most before you search. Sharing a post now sends a link that opens it on cadenceai.fyi, and the website and the Android app now look and work like the iPhone app.
Live Photos in a post or story now move: each one plays its few seconds of motion as a smooth loop in the slideshow, just as it does in the Photos app. Tap LIVE on it before posting to share the still photo instead.
Wraps filmed with the front camera on iPhone 17 came out sideways and zoomed in, because that camera is mounted differently from older ones. Cadence now asks each camera which way is up, and the iPhone 17’s square front camera films a proper upright 9:16 picture. Wraps also no longer come out more zoomed in than the camera showed while filming.
Cadence on Apple TV has an Ambient Mode to leave on while you get ready: a big clock, the class you’re in with a ring showing how long is left, and what’s next. Set how long it takes to get to school and Today shows when to leave, with a reminder when it’s time. Today also shows when school ends and how long until the weekend or your next day off.
The new Focus Timer lets you study with a 25, 45 or 60-minute timer on the TV, and a finished session counts toward your study streak on your iPhone. The You tab shows your streak, points and level next to the days you’ve rated, and you can play the song you picked for a day.
Ask Siri “When does school end today in Cadence?” or “How long until my next class in Cadence?” on your iPhone, Apple TV or HomePod.
There’s now a Cadence app for Apple TV. Connect it from your iPhone in Settings › Apple TV with a 4-digit code, and the TV shows the class you’re in (or the next one) big enough to read from across the room, the next two weeks, the days you’ve rated, and your notifications. It all comes straight from your iPhone over your Wi-Fi, encrypted. No account on the TV, nothing through a server, and messages never show what they say. While Cadence is open on the TV, a banner appears five minutes before each class and when something new arrives on your phone. When it’s closed, the number on its icon counts what you haven’t seen. Removing the TV on your phone clears everything from it.
Ask Siri “What class do I have in Cadence?” on your iPhone, your Apple TV or a HomePod (with Personal Requests on). Settings › Siri & HomePod lists what you can ask.
When you add a song to a wrap you can now choose which part plays: drag along the waveform and it plays from there. The lyrics line up with that part by themselves; if they’re off, tap the line you hear, or nudge them earlier or later.
Cadence now checks for new posts from people you follow, and for likes, comments, replies, reposts, shares, follows and messages, even after you’ve left the app. iOS wakes it every so often to look, and tapping a notification opens the post or chat. While you’re in the app they show as a banner at the top, and the list under the bell is still there after you close Cadence. If notifications are off, the Feed offers to turn them on.
Put a finger on the tab bar and slide: a lens of glass follows it, magnifying the tab underneath, and springs onto the tab you let go on. Taps work the same as before.
The streak and the daily points never moved: a mix-up in the app meant no day was ever saved as a study day. Now any of these keeps your streak going: writing or changing a note, finishing a quiz, practising flashcards, or studying with Study Coach, Onyx or a note’s AI tools. Flashcards earn points too, 2 for each card (up to 60 a day). Tap the progress card on your profile to see what counts and what you’ve done today. Posting still earns points but isn’t a study day.
You can slide your finger along the tab bar to move between tabs, the way iPhone apps do now: the glass follows your finger, and letting go opens that tab.
Messages are back on iPhone, for friends only (people who follow each other). Open them from the paper plane in the Feed or the Message button on a friend’s profile. Send photos, react with ❤️ 😂 😮 😢 🔥 👍, and see when a message has been seen. Every message and photo is checked before it’s sent and again when it arrives, links can only go to school sites, phone numbers, addresses and invite links can’t be sent, and a notification says who wrote to you, never what they said. Report or block from any chat, and lock Messages with Face ID in Settings.
Reposting works like Instagram: tap the repost button to share a post with your followers, who see “you reposted” over it, add a note if you want, and tap again to take it back. Send a post or a wrap to a friend, add a photo to a comment, and follow someone from their face in Wraps.
The Feed has For you, ranked for you on your phone, plus Following, Friends and Latest, with “Why you’re seeing this” and “Not interested” on every post. Swearing can’t be posted any more, including spellings like “fuk” and “sh1t”, and profile pictures are checked like any other photo. Notifications now cover likes, comments, replies, reposts, shares, new followers and new posts from people you follow, and whatever happened while Cadence was closed arrives when you open it; tap one to go straight there.
Tap a class in Schedule to rename it or change its room, teacher or colour. Rate your day has a song of the day now, and its card is story-sized, with your week so far.
Stories have a ··· menu now: delete your own, or report someone else’s. Cadence developers can delete any post, wrap or story from its menu, the same way they already could on the website, and a delete now waits for the server. If it’s refused, the post stays and you’re told why, instead of disappearing and coming back later.
You can post Wraps from the iPhone app now: Create › Wrap. Pick a sound, and it plays while you record, from 15 seconds up to 10 minutes, or choose a video you already have. Cadence finds the song’s lyrics and lines them up with the sound by itself, then shows the line being sung up by the camera while you record, so you can lip-sync. Practise plays the sound with the words before you start. Add text, set how loud your video and the sound are, write a caption, and share. The camera no longer zooms in by itself: recordings keep exactly the frame you saw.
Photos and videos of any size post now. They’re made a sensible size on your phone and sent straight to storage, with a progress bar and a retry, instead of the short request that used to fail with “network” or “too large”. The Feed’s top bar tucks away as you scroll down and comes back when you scroll up.
Everything that was still a web page on iPhone is native now: writing a note (scan pages, add a photo or a PDF), the AI tools for a note, the study coach, Onyx, your timetable and bell times, the school calendar, reporting a post, feedback, your standing and insights. The Feed opens with a free student ID check, and the picture is erased straight after. New: rate your day each evening, tell someone about it, see your days ranked from best to worst, and ask Siri “What classes do I have tomorrow in Cadence?”.
On iPhone a note now opens as a native screen instead of a web page. It fills the screen and closes with an X. Read and Study sit side by side in a switch at the top. Read shows the note set like a page. A PDF opens in Apple’s own reader, so you can select a word in a handout and copy it, look it up in the dictionary or translate it. The old reader drew pages as pictures and never could.
Study shows how much of the note’s deck is waiting, with one button to practise it, and the tools grouped underneath: test yourself, understand, tidy up, homework help. Three rows that were all flashcards under different names are gone. Practice is native too: a card you tap to turn over, and Again, Hard, Good and Easy with the time each would put the card away for. It uses the same deck and the same schedule as before, so nothing you studied is lost and the website agrees.
The bar across the top is Liquid Glass again, in the floating shape you know, with lighter controls. The last time it was native your profile picture vanished in light mode: the page was sending the avatar’s colour as transparent, because the avatar is painted with a gradient. That's fixed at both ends. The bar no longer appears on the sign-in screen, and it steps out of the way of the Study coach, whose close button it used to cover.
A hundred and six rules in the stylesheet set text in capitals, and nearly all of them were ten-pixel labels with wide letter-spacing. The effect on a phone was that the smallest text on every screen was also the loudest, competing with the thing you actually came to read. Those labels now read as ordinary sentences, at a size you can see.
The three shortcuts on Home had two labels each: “VIEW TIMETABLE” stacked above “Schedule”, “SCHOOL POSTS” above “Feed”. Now they have one.
Your profile showed your streak twice: a pill with a dot for each day of the week, and directly beneath it a ring repeating the same number. The ring is gone and the two that remain, points and rank, share the row instead of leaving a hole where the third used to be.
Tapping a note used to spend a quarter of the screen before showing you a word of it: the app’s own bar across the top, then the subject, the title, the date and the card count on four separate lines. The note now fills the screen, the four lines are one, and what you wrote sits on a page of its own instead of running edge to edge on the background.
Under the note there used to be eleven things to tap, five of them flashcards under five different names, all in the same scroll as your writing. They've moved into a Tools sheet. The note keeps one action at the bottom, within reach of your thumb: practise the cards it has, or make them if it has none. Two dead tabs marked “Note” and “Study” were still in the page from a design that was removed. They're gone.
Settings had no close button at all. It has one now, it fills the screen instead of sitting under the app bar, and its seven headings have stopped shouting in capitals. Rows without an icon used to start their label in a different column from rows with one, so the list ran down the page in two ragged edges. They line up now. A section heading that pinned itself to the top was being drawn under the rows it was covering, so text showed through it. The headings scroll with the list now, the way an inset-grouped list on iOS does.
The Study screen’s main button had no visible label in dark mode. The word “Continue” was there the whole time, in exactly the same cream as the button under it, so the screen looked like it had a blank pill at the bottom. The same fault hid “Try again”, “Next question” and “Next card”. One line of theme styling, written for form fields, was reaching the buttons as well and making them inherit the page’s text colour.
The rest of that screen has been rebuilt around it. Each question sits in its own rounded card instead of a thin outlined box. The grades you can aim for no longer wrap, so “100” isn't left alone on a line of its own. The four labels that shouted in capitals now read as sentences. The button sits at the bottom of the screen where your thumb is, instead of below a text box you had to scroll past to reach it.
Practising a note’s cards was a flashcard screen with no card: the question floated in the middle of a flat page with nothing around it. It's a card now, and tapping it turns it over. Again and Easy are tinted so you can tell them apart without reading, and the intervals under them are readable, not grey on grey.
Home used to open with three numbers that never changed: how many subjects you take, how long the cycle is, how many periods there are. In their place, four things that move: how much of today’s studying is done, drawn as a ring; what’s being examined or handed in over the next four weeks, colour-coded; the notes you wrote last; and what’s sitting in your card decks, with a tap to start.
A document note now opens with a cover (the title, how many pages, which file, when you added it) and a scrubber under the page you can drag through the whole document instead of tapping an arrow once per page. You can mark a page, which puts a tick on the scrubber, and write a note against a particular page.
Highlighting and looking a word up aren't here yet. The pages are drawn, not laid out as text, so there's nothing to select, and buttons for them would only pretend. That needs a text layer over every page, and it's the next thing in the reader.
Opening the app used to flash an older-looking version of itself for a second, then redraw. The app is one really big page, and the phone was painting it while still reading it. Now nothing shows until the whole thing is ready, in the theme you last picked.
Scrolling is quicker for five specific reasons: a shader was repainting the bars on every frame forever, a timer was forcing the phone to recompute the layout four times a second, a contrast check was re-reading every element on screen whenever anything changed, nineteen surfaces were running an expensive blur filter, and a cleanup step was firing a fake window resize. The worst frame we could measure went from 30.5 ms to 8 ms.
The buttons in the top bar are smaller, and still just as easy to hit.
The buttons in the top bar were painted white on a pale bar, so in light mode they vanished. They're back, on a bar that's now properly Liquid Glass. Same shape, real material.
Notes used to put six controls between you and your first note. The two view switches and the filter button are gone; search, the category pills and your notes grouped by day are what is left.
Settings looks like Settings: rounded inset groups, dividers that start after the icon, and a search box that finds the row you meant instead of the section it lives in. Liking a post is instant. It used to redraw the whole post just to colour in one heart.
A note used to open as a card floating over a dimmed copy of the list. It fills the screen now, between the top bar and the tab bar, and Settings does the same.
Switching between light and dark left the words in the old theme’s colour, white on a pale page, and it never recovered. A safeguard meant to rescue unreadable text was pinning colours in a way nothing else could override. It lets go on every switch now, and re-checks once the new colours are actually drawn.
A note had two modes and you could only see one at a time. Reading a note and studying it aren't two modes: you read a note, and studying is something you do to it. One screen now: the note, then what the cards are doing (“12 cards · 3 ready · practised 2 days ago”), then one button that knows what comes next (Practise 3 cards, or Make cards from this note if there are none yet) and the rest of the study tools under it.
The Grade field on the Create account form was invisible on a phone, and the form wouldn't submit without it, so nobody could finish making an account. Deleting your account and signing out had also gone missing whenever Settings was filtered to Account, which is the first place anyone would look.
A PDF note was being scaled to fit its height, which put a letter-size page at 261 px across on a phone. It fits the width now and scrolls, the way Files and Books do, and the document runs edge to edge instead of sitting in a frame.
Verifying a school email used to fail with “Check your connection” when your connection was fine. The check just isn't switched on yet. Now it says so, and checks whether you're actually online before blaming you.
PDFs also no longer freeze the app while they open: a security rule was stopping them from being drawn in the background, so all that work happened on the screen’s thread.
Widgets for the Home Screen and Lock Screen showing the next class and the day’s cycle. Study cards that a note writes once and then brings back as you forget them. On iPhone the tab bar is now drawn by the system in Liquid Glass instead of painted by the app, with the selected tab sliding between positions.
Reporting and blocking a post opened a menu that could land off the screen; the notifications panel opened in the top-left corner, half of it off the edge. Both fixed.
Unlocking the Feed with a school email isn't switched on yet, so that check can't pass. The app says so instead of pretending it's your connection. Direct messages are turned off in this version. Anything else you find, tell us. The whole point of a test build is that it breaks.
One document. No build step. Instant everywhere.
Cadence ships as one self-contained web app with a hash router, a glass design system and a Noctic backend behind it.
| Shell | Single HTML document, hash-based routing, no bundler. Loads on any device instantly. |
|---|---|
| Design | Glass: floating pill navigation, 30px blur, hairline borders, edge-to-edge safe areas. |
| Data | Noctic cloud backend with row-level security; media through signed uploads so posts sync across devices. |
| AI | Every model call goes through a server function. No key ever reaches the browser. |
| Identity | Regular Cadence accounts, or Continue with Noctic for one-tap sign-in. |
| Delivery | Installable PWA, Android APK and iOS home-screen app. |
Cadence runs on whatever a student already has, which is usually a four-year-old phone with no storage left. Everything below comes from that.
One document means the app works on a bad connection in the time it takes a native app to show its splash screen.
Nothing to review, nothing to approve. A fix reaches every student the day it's written, which matters a lot when what broke is the timetable in September.
No server, no plugin, no management profile. A school can be up and running by the end of lunch, not the end of a procurement cycle.
The interface is built to stay smooth on the kind of hardware a flagship demo ignores. Nobody should need a new phone to read an announcement.
| Offline | Your calendar and notes stay readable offline. Anything new you write waits and syncs later. |
|---|---|
| Updates | Applied next time you open it. There's no version to chase and nobody stuck on last term's build. |
| Notifications | Quiet hours are on by default for student accounts, and channels you're not in can't reach you. |
| Appearance | Light and dark follow your system, or you can set it yourself and each device remembers. |
| Accessibility | Keyboard operation, screen-reader labels and reduced motion throughout. Current gaps are listed on Accessibility. |
| Model access | Every call goes through our server and is rate limited. No key is ever in a browser, on any page, in any build. |
The sexyest design language out there.
Every page starts at night, blue light over navy, and turns to day as you read. Bold type holds the structure, one blue does the pointing, and nothing gets an outline.
Real refraction, not a blur. Every pixel behind the panel is bent through a bevelled glass surface with Snell's law and per-channel dispersion at the rim, the same optics we tune in Shiba and Cadence. It behaves like the material it's drawn as, too: pulled along after your hand instead of pinned to it, stretched in the direction you throw it, ringing where you touch it, and fusing with any drop you bring it near.
The palette is one sky. Navy at the top of the page, royal blue at the horizon, then ice and white. Royal blue is the only colour that ever asks you to do something.
SF Pro everywhere. Headlines are weight 800 with tight spacing, and the text underneath stays regular, so weight alone tells you what matters.
Make it intuitive.
One account for all of it.
Works ahead of your school year.
Consistency you can see.
Sign in once and every Noctic app knows it's you. The text that explains things stays plain, so the headline above it can do the work.
Everything you press is a pill. Everything that holds something is a rounded, tinted surface.
The bars are clear liquid glass. You see straight through them, the edges bend what’s underneath like a lens, and the rim catches the light. Never solid, over the dusk or over the day.
One solid blue pill for the action that matters, and a soft blue one next to it for the next best.
Fields are tinted wells that turn white with a blue ring when you type in them.
Panels lift off the page with a long, low blue shadow, not an edge.
The light at the top of every page is real. Particles drift, form the Noctic droplet on the home page, move away from your pointer, and spread out as you scroll into the day. An animation has to explain a change. If it's only there to look expensive, it gets deleted.
Reduced motion matters just as much. Every transition has a still version that means the same thing, used automatically when your device asks for less movement, and the particles settle into a still constellation. Settings → Appearance overrides it either way.
Auth, billing, AI gateway, storage and security, shared across every Noctic product.
Build on the platform and the second your app goes live you get years of work on rate limiting, row-level security, scoped consent and model routing for free.
Authorization code with scoped consent. Tokens can be revoked from the account page whenever.
OAuth 2.0, scoped consent, revocable tokens and verified profile propagation.
One subscription follows the account. If a payment fails, access is cut right away.
Rate-limited, model-routed, keyless access to the AI layer from any app.
Signed uploads, bucket-level RLS and cross-device media sync.
Column-level grants, security definer functions and continuous dependency scanning.
Serverless functions with mandatory JWT validation and automatic abuse limits.
These are the parts nobody plans for and everybody ends up having to build.
| Sign-in that already works | No password reset flow to write, no sessions to get wrong, no verification email to send. The account exists before your app does. |
|---|---|
| A plan that follows the user | Someone who upgrades in Shiba is upgraded in your app on the very next request. You never have to build billing to find out. |
| Model access without keys | Requests go through us and get routed and rate limited on our side. There's no key in your client to leak and no provider account for you to manage. |
| Consent you didn't design | People see exactly which app is asking for what, approve it scope by scope, and revoke it from one page. Revoking is instant, not eventually. |
| Abuse limits by default | Rate limiting comes switched on, not as something you add after your first bad week. |
| Age-aware behaviour | Assistant tiers follow the verified age on the account, so an app students use behaves right without you writing the rules. |
Ask for the scopes you use. Asking for more than your app needs gets refused at review.
When someone revokes access, tokens stop working right away. There's no grace window to abuse.
Client secrets go on your server. An app that ships one in front-end code gets suspended.
The authorisation flow needs it, and we reject callbacks that come back without a matching value.
Profile data from sign-in can't be sold, brokered or merged into someone else's graph.
Apps that reach student accounts have to follow the extra rules in Children and Students.
Name it, set a redirect URL, and pick the scopes you actually need. Register.
Authorisation code with state, exchanged on your server. Ready-made examples are in Code examples.
The developer tools let you inspect tokens, check webhook signatures and see what a real consent screen gives back.
Scopes past basic identity get reviewed before release. If something's missing we tell you, we don't just quietly reject it.
One account, unlimited services. Great on their own. Even better together.
Start a note in Cadence, finish it with Shiba. Your context comes with you.
Noctic Plus and Noctic Pro work across every Noctic product you use.
Your Noctic Account carries your name, avatar, verification and permissions.
Consent screens show exactly which app is asking for what, every time.
| Noctic Account | The identity layer. Hands out authorization codes and tokens to every app. |
|---|---|
| Shiba | The AI layer. Runs Cadence AI and any app you connect. |
| Cadence | The social layer. Signs in with Noctic and uses Shiba for its AI. |
| Your app | Register it, ask for scopes, and get the whole stack. |
Everyone says their stuff integrates and you never actually notice. These are the exact moments where using two of our products beats using one.
| Verification travels | A blue check you earn in Cadence shows up anywhere your account does, with no second ID check. |
|---|---|
| One upgrade, three products | Upgrade in Shiba and your limits in Cadence go up on the next request, not next billing cycle. |
| One consent screen | Every app that wants something asks in the same words, in the same spot, with the same revoke button under it. |
| One deletion | Delete the account and it's gone from everything. No product is left holding a copy. |
Cadence is worth using without a Shiba plan. Shiba is worth using without a school. Nothing here holds something else hostage.
Export what you've written, in a format you can open, without asking anyone.
Any connected app can be cut off from one page, and access stops the second you do it.
Continue with Noctic is open to apps that aren't ours. The ecosystem isn't just a closed set of our own products. Developers.
Manage your Noctic Account
Don't have a Noctic Account? Create yours now
We sent a six-digit code to your address. Enter it to finish signing in.
No code yet? Check your spam folder. The same email carries a sign-in link that works just as well if you'd rather tap it. Wrong address? Contact support.
The code couldn't be sent, so nothing's coming. Nobody's signed in and nothing about the account changed. Tell support and we'll fix the delivery.
There were five failed sign-in attempts on this account, so we locked it and emailed an unlock link to the address on file. Open that email and follow the link to confirm it's yours.
The lock lifts by itself in 15:00, or right away once you use the link.
Continue to
This app is asking to use your Noctic Account.
Connecting as this account. Use Switch if you have more than one. You'll be asked for the password and an emailed code before anything is connected. You can revoke access any time from your account.
Sign in to manage your details, your devices and the apps you've connected.
Your account
One account for Shiba, Cadence and every app you connect with Continue with Noctic.
About you
The details on your account. Your name is what apps see when you approve the profile scope. The rest stays with us.
Your date of birth turns on age-appropriate features, so you can't edit it here. Contact support if it's wrong. Language and text size are in Settings.
Protection
Your email address is your account name. Every sign-in also needs a code we email you, so a password alone is never enough.
Passwords are checked against known breaches in your browser before we accept them, and the password itself never leaves this device. More in Security.
Access
Every device signed in to this account and its recent sign-ins. Sign out any one of them, all the others, or every device including this one, which also revokes the tokens held by the Noctic apps.
Connections
Apps you've approved with Continue with Noctic, what each one got, and a button that ends it. Revoking works right away.
Apps only ever see the scopes you approved. Your date of birth, plan history, moderation records, private notes and assistant conversations aren't available through any scope. How the ecosystem works.
Subscription
What this account is on right now. Plans are bought inside the product you're using, and cancelling kicks in at the end of the period you've paid for.
Card details are handled by Stripe and never reach our systems. Contact us about invoicing or an education agreement.
Your data
Take a copy whenever you want, or close the account for good. Neither costs anything, and neither needs a chat with us first.
Deletion removes your profile, posts, notes and generated content across everything. It can't be undone and it's done within thirty days. How privacy works.
Apps ask for identity, profile, email or API access, and you approve each one.
Remove any app's access whenever from your account settings.
Birthdays, plan history and moderation records are never shared with apps.
Verification you earn in one product counts across all of them.
Four scopes, no bundles, and nothing granted without you saying so.
| identity | That you're a real signed-in Noctic user, and a stable identifier for you. Nothing else. |
|---|---|
| profile | Display name, avatar and verification badge. Stuff other people can already see. |
| Your email address, so the app can contact you. Asked for separately and you can say no to just that. | |
| api | Permission to make model requests for you, counted against your plan and shown in your usage. |
Your date of birth, plan history, moderation records, private notes, direct messages and assistant conversations aren't available through any scope. There's no partner tier that unlocks them.
A list of every app with access, what it asked for, and when you approved it.
One button ends access right away. No grace period, no token that keeps working for an hour.
New passwords are checked against known breach lists before they're accepted, and the password never leaves your browser.
Take a copy of everything, or delete the account and its content across every product, from Settings.
Sign-in, keys, Cadence, webhooks and the Program, all on one page.
Start
Noctic is one account across Shiba, Cadence and everything else we make, and building on Noctic means building on that account. There are three ways in.
| Address | What it's for |
|---|---|
developer.noctic.ink | This site: the docs, the platform, keys, tools, examples, Kick Start, the Program, Shouts and events. |
noctic.ink | The Noctic Account itself: sign-in, profile, devices and connected apps. |
shibaai.dev/oauth/authorize | Where your app sends people to sign in and approve it. |
oijfjxqcavxiexnyszbu.supabase.co/functions/v1 | The token and userinfo endpoints. |
api.noctic.ink | Class Sync at /class-sync, the Cadence Teacher API at /teacher and the Parable API at /parable. |
Start
From nothing to a signed-in user. You need a Noctic Account (if you're reading this signed in, you have one) and a server that can keep a secret.
state.
?code=shc_…&state=….
Check state is the value you sent before you do anything else.
Try every step against the live endpoints first. The URL builder, token exchange tester and userinfo explorer run in your browser, as you. Put your client ID into Code examples and every sample rewrites itself with it.
Start
Everything on this site happens as your Noctic Account, so every key can be traced back to whoever made it and revoked by them. No separate developer login and nothing extra to set up.
noctic.ink and developer.noctic.ink are one family of sites and share one session. Sign in on either and the other knows straight away. Come here from a browser where you're already signed in to Noctic and there's nothing to press. Signing out works the same way: sign out of one and you're signed out of both, on this device.
Not signed in anywhere? Sign in takes you to noctic.ink for your password and the emailed code, then straight back to the page you were on, signed in.
Picked not to stay signed in? Then the session lasts until you close the browser, on both sites, and isn't saved on the machine.
Everyone who signs in to your app can see it in their Noctic Account with the scopes it has, and disconnect it whenever. When they do, its tokens stop working right away, no grace period. Treat a sudden 401 as a disconnect and send them back through sign-in, don't keep retrying.
Sign in with Noctic
Sign in with Noctic is OAuth 2.0 with the authorisation code grant. Your app never sees a password. People sign in on Noctic, approve the scopes your app wants, and come back to you with a short-lived code your server trades for tokens.
code and your state.| Endpoint | Address |
|---|---|
| GETAuthorise | https://shibaai.dev/oauth/authorize |
| POSTToken | https://oijfjxqcavxiexnyszbu.supabase.co/functions/v1/oauth-token |
| GETUserinfo | https://oijfjxqcavxiexnyszbu.supabase.co/functions/v1/oauth-userinfo |
A plain link or a redirect. Every parameter goes in the query string.
| Parameter | Required | What it is |
|---|---|---|
client_id | Yes | Your app's client ID. |
redirect_uri | Yes | Where to send the person back. Must exactly match one registered for the app. |
response_type | Yes | Always code. |
scope | No | Space-separated. identity profile email covers most apps. |
state | Recommended | A random value you save and check on the way back. It's what stops a fake callback. |
code_challenge | Public clients | The PKCE challenge, for apps with no secret. |
code_challenge_method | With a challenge | S256. |
From your server, POST JSON to the token endpoint. The code begins shc_, expires sixty
seconds after it's issued and works once, and redirect_uri has to be byte-for-byte the
one used to get it. Nothing gets sent there this time. It's part of the proof the exchange
belongs to that sign-in.
| Field | What it is |
|---|---|
access_token | A signed JWT. Send it as Authorization: Bearer …. Valid for one hour. |
refresh_token | Gets a new access token without bothering the person again. Store it on your server, encrypted, like the real credential it is. |
expires_in | Seconds until the access token expires. |
scope | What was actually granted, which can be less than what you asked for. Read it, don't assume. |
Access tokens last an hour. Refresh shortly before one expires instead of waiting for a request to fail, and keep the new refresh token that comes back with the new access token.
Single-page and mobile apps ship their code to the people using them, so they can't keep a client secret. They send a code_challenge with the authorisation request instead, and prove it with the matching code_verifier when they exchange the code. The PKCE generator makes a pair to test with.
One GET with the access token. The response is shaped by what the person approved, not what you asked for, and a field you were refused is missing, not empty. Check before you read it. Call it once after sign-in and store what you need, because it's rate-limited per token.
| Scope | What it adds |
|---|---|
identity | The stable Noctic user ID, sub. Proof of a real account and nothing more. |
profile | Username, display name, picture, plan and verified status. What other people already see. |
email | email and email_verified. Refusable on its own without refusing the rest. |
api | Model requests on the person's behalf, drawn from their plan. No extra profile fields. Granted on request, so ask us. |
Date of birth, plan history, moderation records, private notes, direct messages and assistant conversations aren't available through any scope, and there's no partner tier that unlocks them.
| You see | What it means |
|---|---|
invalid_grant | The code expired, was already used, or was issued for a different client or redirect URI. Codes are fragile on purpose. Just start the flow again. |
invalid_client | The client ID and secret don't go together. Roll the secret on Keys if you lost it. |
401 from userinfo | The token expired or the person disconnected your app. Refresh once. If that fails too, send them back through sign-in. |
403 from userinfo | The token is fine but doesn't have the scope. Ask for it in the authorisation URL, because a scope can't be widened afterwards. |
Build
An app is a name, a homepage and one or more redirect URIs, and it has two credentials.
| Credential | How to treat it |
|---|---|
client_id | Public. It goes in the authorisation URL and names your app on the consent screen. |
client_secret | Private. Shown once, when the app is made or the secret is rolled, and never again. It belongs in an environment variable or a secret manager on your server. Not in a repository, a front-end bundle or a screenshot. |
http://localhost while you develop.localhost may use http://. Everything else should be https://.redirect_uri in a request has to match a registered one exactly: scheme, host, port and path.A lost secret can't be recovered, only replaced. Rolling it on Keys issues a new one, and the old one stops working the moment it does, so have the deploy ready before you roll. Sign out all users on the same page ends every session your app holds, on every device.
Build
Every webhook we send carries its signature in the X-Noctic-Signature header:
sha256= and then the HMAC-SHA256 of the raw request body, keyed with your signing secret,
in lowercase hex. A request whose signature doesn't match didn't come from us. Drop it without doing
anything else.
Verify, answer with a 2xx, then do the work. Slow handlers cause retries and retries cause duplicates, so make the handler idempotent and key it on the event's ID. The webhook signature tool makes the exact header for a payload so you can test your check.
Build
Glass is the design language behind everything Noctic makes: system type doing the hierarchy, one blue accent, hairline borders, blurred surfaces over a single background, two corner radii and one motion curve. The design kit packs it as a skill (the rules, every token in light and dark, the components and the way we write) so Claude can build your interface in the Noctic look, or you can read it and build it by hand.
Unzip it where your tools look for skills. Then ask for the Noctic look, like “build the settings page in the Noctic style”, and Claude reads the kit before it writes anything.
| File | What it holds |
|---|---|
SKILL.md | The four layers, the rules we never break, type, colour and a checklist to go through before anything ships. |
assets/glass.css | Every token in light and dark, and the components built from them: the glass bar, stages, pill buttons, cards, fields, status pills, code blocks and sheets. Drop it into any site. No build step. |
assets/starter.html | A whole page made from nothing but glass.css, with an appearance switch, to copy from. |
references/components.md | Markup, sizes and states for each component, and the motion table. |
references/voice.md | How we write buttons, headlines, errors and empty states. |
cubic-bezier(.16, 1, .3, 1). With reduced motion it fades instead, and nothing goes missing.The look, not the brand. Use Glass freely in your own product. The Noctic mark, the name and our product names stay ours, so keep them out of your logo, app name and copy. The one place they belong in your app is the button above. The brand guidelines have the detail.
Cadence
A class website, read into a class. Your site sends a manifest of what it publishes (decks, notes, handouts, links) and Cadence fetches each item, pulls the text out of it and files it under the class the key belongs to. Everything a student sees comes from something the teacher posted, and everything Cadence AI says about the subject can be traced back to it.
| Key | Where it lives, and what it can do |
|---|---|
ntk_pub_… | Publishable. It sits in a script tag right on the class website, so it's locked to the class's own sites and can do one thing: offer material for its class. It can't read anything back or see a student. |
ntk_sec_… | Secret. Server-side only, for an LMS or CMS plugin. Reads and withdraws material as well as posting it, and is the only key the Teacher API accepts. |
Get both from Class keys. We store a hash of each key, not the key, so a lost one is replaced, not recovered.
Base URL https://api.noctic.ink/class-sync. Send the key as a bearer token, or in X-Noctic-Key.
| Call | What it does |
|---|---|
POST/verify | Checks a key and returns its class. Changes nothing. Call it first and show the class name, so a key pasted into the wrong site is obvious. |
POST/materials | The manifest, up to 500 items. Items already on file are matched on id and updated, so posting the same list on every page load is totally fine. |
GET/materials | What is held for the class and what happened to each item. Secret key only. |
DELETE/materials?id=… | Withdraws one item from students and from Cadence AI. Secret key only. |
Only id and url are required on an item. title, kind, topic, checksum and posted_at make it easier to find and cheaper to sync. A top-level complete: true says the manifest is the whole class, so anything missing from it gets withdrawn. Leave it off when you sync one page at a time. The full Class Sync reference has every field, the file types Cadence reads, and the drop-in script that does all of this for a normal class page.
Cadence
Class Sync handles material. The Teacher API is everything else about a class: its name and term, the sites allowed to speak for it, and the roster. It's for schools whose register lives in an SIS or an overnight CSV instead of a web page.
Secret key only, and from a server only. It refuses ntk_pub_ outright and sends no
CORS headers, so a browser will never hand a roster to a script, even if somebody pastes a secret key
into a page.
Base URL https://api.noctic.ink/teacher, the same authentication as Class Sync, and a
budget of 60 calls a minute shared with Class Sync, because it belongs to the key. There's no class ID
to pass: the key is the class.
| Call | What it does |
|---|---|
GET/class | The class, its allowed origins, whether sync is live, and counts of the roster and material. |
PATCH/class | name, subject, school, term, site_origins, live. Fields left out, or sent empty, are left alone, except site_origins: [], which clears them. |
GET/students | The roster. ?include=removed adds people taken off it. |
POST/students | {"handles": ["alice", "@bob"]}, up to 200 at a time. A string is split on commas and whitespace. |
DELETE/students?handle=… | Takes one person off, immediately. Removing somebody already off isn't an error. |
Each handle comes back active (they have an account and are in the class), invited (nobody has that handle yet, and it turns on by itself when they sign up), already or invalid. If a whole export comes back as invited, it usually means the export has the wrong column in it. The full Teacher API reference walks through a full nightly sync.
Parable
Parable 1 is our own model, and this is how you put it in your app. One endpoint, one model, and you pick how hard it thinks. It speaks the same chat completions shape most SDKs already know, so pointing an existing client at it is usually just a new base URL and a new key.
Keys go on a server. A Parable key spends your token budget, so it never goes in a page or an app bundle. The API sends no CORS headers, so a browser won't get an answer back even if a key ends up in one.
Base URL https://api.noctic.ink/parable. Send your key as a bearer token. Keys start with pk_live_ and you make them under Parable keys. We only keep a fingerprint of each one, so a lost key gets replaced, not recovered.
| Call | What it does |
|---|---|
POST/v1/chat/completions | Send messages and get an answer. model is parable-1, effort picks how long it thinks, and stream: true sends it back as server sent events. |
GET/v1/models | The model and its effort levels. Doesn't need a key. |
GET/v1/status | Whether Parable 1 has somewhere to run right now, and how many tokens your key has left today. |
Each message has a role (system, user, assistant or tool) and text content. Images and voice work in Shiba but aren't on the API yet, so a message with an image in it gets a clear 400 instead of being quietly dropped. When the model shows its working, it comes back in message.reasoning next to the answer, not hidden.
There's one model, not a menu of them. Effort is how long it gets to think before it answers, and it's the main thing that decides cost and speed. Leave it out and you get medium.
| Effort | Good for |
|---|---|
economic | Labels, short rewrites, yes or no. Fastest and cheapest. |
low | Everyday questions and quick summaries. |
medium | The default. Most chat, most writing. |
high | Maths, code and anything with several steps. |
plus | Long problems where it should plan and check itself. |
ultra | The hardest work. Slowest, and it uses the most tokens by far. |
Parable 1 can run on our servers or right in a browser on a Noctic machine, and the API splits the work between the two. Lighter effort usually runs locally and heavier effort goes to a server. If one side is busy or offline the other one picks it up, so your code never has to care. Every reply says which one answered in noctic.route, either server or local.
A prompt sent to a local machine only waits in the queue until it's answered, and then it's wiped. Nothing you send is used to train the model.
Each key gets 60 requests a minute and 200,000 tokens a day to start with. Both are counted per key, and Parable keys shows where you're at.
| Code | Meaning |
|---|---|
400 invalid_request | Something in the body is wrong. The message says which field and why. |
401 no_key | No bearer token was sent. |
401 unknown_key | The key was mistyped or revoked. Make a new one under Parable keys. |
429 rate_limited | Over 60 requests this minute. Wait for Retry-After. |
429 daily_budget_spent | The key has used today's tokens. It resets at midnight UTC, and Retry-After says how long that is. |
503 no_capacity | Nowhere to run it right this second. Nothing was charged. Try again after Retry-After. |
504 timed_out | It didn't finish in time. Nothing was charged for the answer. Try again or use a lower effort. |
Reference
Every refusal says what it is and, when it can, what to change. These are the ones worth knowing before you ship.
| Limit | Applies to | What happens |
|---|---|---|
| 60 calls a minute | Each class key, across Class Sync and the Teacher API | 429 rate_limited, with Retry-After in seconds |
| 500 items, 512 KB | One Class Sync manifest | 413 too_many_items; split it |
| 200 handles | One POST /students | 413 too_many_handles; send batches |
| 200 MB | One file read by Class Sync | The item is marked skipped with a reason |
| 60 seconds | An authorisation code | invalid_grant on exchange |
| 1 hour | An access token | 401; refresh it |
| Code | From | Meaning |
|---|---|---|
401 no_key | Teacher API | No bearer token and no X-Noctic-Key. |
401 malformed_key | Teacher API | That isn't the shape of a class key. |
401 unknown_key | Both | The key isn't active. It was rolled, revoked or mistyped. Make a new one under Class keys. |
403 origin_not_allowed | Class Sync | A publishable key used from a site the class doesn't list. Usually www, staging, or http against a registered https. |
403 read_needs_secret_key | Class Sync | A publishable key tried to read or withdraw. That's on purpose. |
403 needs_secret_key | Teacher API | You sent the publishable key. |
405 no_browser | Teacher API | A CORS preflight arrived. Call it from a server. |
409 class_paused | Class Sync | The class exists but sync is off. Turn it on under Class keys. |
413 too_many_items | Class Sync | Over 500 items or 512 KB in one manifest. |
413 too_many_handles | Teacher API | Over 200 handles in one call. |
429 rate_limited | Both | Over 60 calls in a minute on one key. Wait for Retry-After. |
Malformed items inside an otherwise good Class Sync manifest are counted in ignored and
rejected instead of failing the call. One bad date shouldn't cost a teacher a term of
material.
Reference
Before you ship, every one of these should be true.
ntk_sec_ key live on a server, in an environment variable or a secret manager.state, and the callback refuses one that doesn't match.S256, and never a client secret.scope that comes back.401 sends the person back through sign-in instead of retrying in a loop.Found a vulnerability in Noctic itself? Tell us through Security, not an issue anybody can read.
Reference
Tools that run in your browser, as you, against the live endpoints. Nothing you type leaves your device except the call you choose to make.
Community
Everything above is free and open to any Noctic Account. The Program is for the stuff that needs a name attached: a verified publisher on the consent screen instead of "an unverified developer", a listing in the Ecosystem, higher limits, early builds, hardware and the labs. Membership is free.
| Membership | For |
|---|---|
| Individual | One person publishing under their own legal name. |
| Organisation | A company publishing under its registered name. Needs an entity number, an email at the company's own domain, and someone who can sign for it. |
| Education | A school, college or university. No entity number asked for. |
| Enterprise, in-house | Apps used inside your own organisation and never listed publicly. |
| Status | What it means |
|---|---|
| Draft | Being filled in. Nobody but you can see it. |
| Needs a change | Sent back, by the automatic check or by the team. The note says what to change. |
| Waiting for review | Complete, and with the Developer team. |
| Member | Accepted. The membership is active. |
| Declined | Not accepted this time. The note says why, and Support is the way to ask again. |
Community
Developer Shouts are short posts from the Noctic Developer team: releases, heads-ups about changes, events and anything else worth a developer's attention. They're on developer.noctic.ink/shouts, and the latest three sit at the top of the developer home. Pinned shouts stay first while they're still true, and anything that changes how an API behaves gets written into these docs too.
Events (sessions, labs and meetups) are listed on Events once they're scheduled, and you take a seat there with the account you already have. Program members get the labs first.
Community
No. Registering apps, the identity, profile and email scopes,
Class Sync, the Teacher API and the Program are free. Only the api scope draws on a
plan, and it draws on the user's.
Usually because you're not signed in on noctic.ink in this browser either, or you signed out of one of them, which signs you out of both. Press Sign in and you'll come straight back here. If you're signed in on noctic.ink and still see this, reload the page once and tell us if it keeps happening.
It can't be recovered, only replaced. Roll it on Keys; the old one stops working as the new one is issued.
Almost always the body. A framework that parsed the JSON already threw away the bytes you need. Use its raw-body option, and check the secret for a trailing newline.
The automatic check answers the second you submit. A complete application then waits for the Developer team, and the decision shows up on your membership page as soon as it's made.
Build the link your app sends people to.
Not sure what a redirect URI is? Read the plain-English guide in the Field Guide.
Five parameters, and only the first two are required. Everything the builder makes is a plain URL. You can read it, paste it into a browser, and see exactly what your users will see.
| client_id | Your app's public ID, given to you when you register it. Fine to ship in a mobile app or a front-end bundle. It says who you are but doesn't prove anything on its own. |
|---|---|
| redirect_uri | Where we send people once they've decided. It has to match one of the URIs registered for the app, character for character, including the scheme, any port and any trailing slash. |
| scope | A space-separated list from identity, profile, email and api. Ask for as little as you need. Short lists get approved way more often than long ones. |
| state | Any random string. We hand it back untouched on the redirect, and checking it against what you sent is what stops cross-site request forgery. Optional in the spec, expected in real life. |
| code_challenge | The S256 hash of a secret your app keeps for the length of the flow. Required for anything that can't hold a client secret: single-page apps, mobile apps, desktop apps. |
Your app opens the URL this page builds. Nothing has been granted yet.
We show who's asking and exactly which scopes. They can approve or refuse each request.
Your redirect URI receives a one-time code and the state you sent, as query parameters.
Your server swaps that code for tokens. The code dies on first use, or after sixty seconds.
The URI in the link isn't one of the URIs saved against the app. The comparison is exact, so
http versus https, a stray trailing slash or a missing port will all fail.
Fix it under API keys.
The client ID doesn't exist, or the app has been deleted. Copy it again from API keys instead of typing it out.
You asked for a scope the app isn't allowed to request. Scopes are set at registration, so widen them there first, then rebuild the link.
They said no, or closed the window. That's a normal outcome, not an error. Handle it by explaining what your app can't do without the grant, and offer the link again.
Yes. This page is for checking your parameters while you build. In production, build the URL in your own code so the state and code challenge are fresh every time.
Create a verifier and its S256 challenge for public clients.
A client secret only works if it can stay secret. A single-page app ships its source to the browser, and a mobile app can be unpacked, so neither can keep one. PKCE swaps the secret for a value made fresh for each sign-in and used exactly once.
A long random string, 43 to 128 characters. Keep it in memory or session storage for the flow.
SHA-256, then base64url with the padding stripped. That result is the challenge.
It rides along on the authorisation URL. The verifier itself never leaves your app.
You send the verifier when exchanging the code. We hash it and compare. No match, no tokens.
Someone who steals the authorisation code still can't use it, because they never saw the verifier that has to go with it.
Yes. Reusing a verifier defeats the whole point. Make one when the user starts signing in and throw it away once you have the tokens.
We only accept S256. The plain method sends the verifier in the clear on the first request, which leaves it in browser history, referrer headers and server logs.
Session storage is the usual answer for a web app, so it dies with the tab. Stay away from local storage, and anywhere another site or another app on the device could read it.
Nine times out of ten, the verifier sent at the end doesn't match the challenge sent at the start. Usually a new one was generated in between, or a page reload wiped it.
The values here come from your browser's own crypto.getRandomValues, so they're properly random. Use them to test the flow, and let your app make its own in production.
Swap an authorisation code for tokens against the live endpoint.
This calls the live endpoint, so you need a real code from a real redirect. Codes are single-use and short-lived. If you've already exchanged one, start the flow again to get another.
Begins shc_, arrives as the code query parameter on your redirect URI, and
expires sixty seconds after it's issued.
Byte-for-byte the one used to get the code. Nothing gets sent there this time. It's part of what proves the exchange belongs to that authorisation.
For confidential clients only. If your app is a single-page or mobile app, leave it empty and send a PKCE verifier instead.
In production this call goes on your backend. A secret that reaches the browser is a published secret, no matter what the bundler promises.
| access_token | A signed JWT. Send it as Authorization: Bearer …. Valid for one hour. |
|---|---|
| refresh_token | Trades for a new access token without bothering the user again. Store it on your server, encrypted, and treat it like a real credential. |
| expires_in | Seconds until the access token dies. Refresh before then, not after a request already failed. |
| scope | What was actually granted, which may be narrower than what you asked for. Read it, don't assume it. |
The code expired, was already used, or was issued for a different client or redirect URI. Codes are fragile on purpose. Run the flow again.
The client ID and secret don't match a registered app. Roll the secret from API keys if you have lost it.
Both stop working right away. There's no grace period. Treat a sudden 401 as a revocation and send the user back through authorisation, don't retry.
Decode a JWT locally and see when it expires.
This page splits the token on its dots and base64-decodes the middle part. That's all decoding is. Nothing here checks the signature, and nothing here leaves your browser. Anyone can read a JWT. Only your server, holding the right key, can tell whether to believe it.
The signature against our published keys, then exp, then iss, then aud against your own client ID. In that order, and reject on the first one that fails.
The alg field comes from whoever made the token. Pin the algorithm you expect in your own code, don't read it back out of the token.
A JWT is signed, not encrypted. Anyone holding the token can read every claim in it, including the browser it went through.
Expiry is checked against your server's clock. A drifting clock makes tokens look expired the second they arrive, or lets them hang around too long.
| sub | The stable Noctic user ID. This is the value to store against your own records. |
|---|---|
| iss | Who issued the token. Always a Noctic issuer for tokens from our endpoints. |
| aud | Who it's for. Should be your client ID; if it isn't, the token isn't yours to accept. |
| exp | Expiry as a Unix timestamp. One hour after issue for an access token. |
| iat | When it was issued. Useful for spotting a clock-skew problem between machines. |
| scope | What the user actually granted. Enforce this on every request, not only at sign-in. |
Present only when the email scope was granted and approved. |
A user ID is stable for the life of the account. An email address isn't. People change them. Key your
database on sub.
Call the profile endpoint with a token and read the response.
The response depends on what the user approved, not what you asked for. A field you were refused is missing, not empty, so check it's there before you read it.
| identity | sub only. Proof of a real signed-in account and a stable identifier for it. |
|---|---|
| profile | name, picture and verified. What other users already see. |
email and email_verified. Requested separately, and refusable on its own
without refusing the rest.
|
|
| api | No extra profile fields. It grants model requests against the user's plan, nothing more. |
Date of birth, plan history, moderation records, private notes, direct messages and assistant conversations aren't available through any scope, and there's no partner tier that unlocks them.
No. Call it once after sign-in, store what you need, and refresh it now and then (daily is plenty). The endpoint is rate-limited per token, and a chatty integration will hit that limit.
It expired or got revoked. Refresh it once. If the refresh fails too, the user pulled access and needs to authorise again.
The token is fine but doesn't have the scope. Ask for it in the authorisation URL. You can't widen a scope after the fact.
Either the profile scope was refused, or the account has no display name set. Fall back
to something neutral instead of showing an empty space.
Cache it, but re-fetch every so often. Avatar URLs change when someone updates their picture, and the old one stops working.
Sign a payload with HMAC-SHA256 to verify deliveries.
Every webhook we send has an HMAC-SHA256 signature over the raw request body, made with your signing secret. Recompute it on your end and compare. If they don't match, the request didn't come from us and you should drop it without doing anything else.
The exact bytes received, before any JSON parsing. Re-serialising changes them and breaks the hash.
SHA-256, keyed with your signing secret, rendered as lowercase hex.
Use your language's timing-safe comparison. A plain equality check leaks the answer a byte at a time.
Reject anything older than five minutes, so a captured delivery can't be replayed at you later.
Almost always the body. A framework that parses JSON and hands you an object has already thrown away the bytes you need. Use your framework's raw-body option.
Check for a proxy that rewrites the body, and check for a trailing newline your shell added when you copied the secret into an environment variable.
Verify, acknowledge with a 2xx, then process. Slow handlers cause retries, and retries cause duplicates. Make your handler idempotent and key it on the event ID.
Roll it from API keys. Deliveries signed with the old secret stop verifying at once, so deploy the new value before you roll.
It's signed, not encrypted. The transport is HTTPS. The signature proves who sent it and that nothing changed on the way, but it doesn't hide the contents from anything terminating your TLS.
One form. Credentials in seconds.
Your client secret is shown once, right after the app is created. Store it somewhere safe. It's never shown again.
Registering takes about a minute, and nothing here is permanent. You can change every field later from API keys.
It shows on the consent screen, above the list of scopes. Use the name your users actually know the product by, not an internal codename.
Somewhere a curious user can go to figure out who you are. Optional, but apps without one get approved less often when they ask for more than identity.
Where we send people after they decide. Add every environment you'll actually use: production,
staging and http://localhost for development.
An environment variable or a secret manager. Not a repo, not a front-end bundle, and definitely not a screenshot in a group chat.
A client ID you can publish, and a client secret shown exactly once on this page.
Where you can edit URIs, roll the secret, grab the button markup or delete the app completely.
Authorisation URL, then code exchange. Our tools check each step against the live endpoints.
New apps start with identity, profile and email. Ask us if you need api.
No. Registering an app and using identity, profile and email
is free and always will be. Only the api scope draws on a plan, and it draws on the
user's, not yours.
Yes, and you should. Separate credentials for production and development mean a leaked development secret can't touch real users.
It can't be recovered, only replaced. Roll it from API keys. The old one stops working the second the new one is issued, so deploy carefully.
Not for the three default scopes. Apps requesting api are looked at by a person first, so
write to support with what you're building and roughly what volume you expect.
Profile data obtained through sign-in may not be sold, brokered or merged into a third-party graph. The full terms are in the legal centre.
Everything you hold, in one place.
Three kinds live here. App keys are for signing people in with Noctic. Class keys are for teachers pushing their own material into Cadence. Parable keys are for calling Parable 1 from your own app. They have nothing to do with each other except they're all yours, and it was easier to find them in one place.
A key you paste into the class website you already have. Every deck, set of notes and handout you post there shows up in Cadence under that class, and Cadence AI answers your students using the slide you actually taught from. You say who those students are, and only they can read it.
A Parable key lets your server call Parable 1. It spends your token budget, so it goes in an environment variable on a server, never in a page or an app.
Your new key. Copy it now. You won't see it again, we only keep a fingerprint of it.
Pick an app above, then copy the markup. It renders the official Continue with Noctic™ element and sends people through the authorisation flow.
Preview
Markup
Choose an app above to generate your snippet.
A client ID is public on purpose. A client secret is a password for your app, and everything that goes for a password goes for it too.
If a secret is in a browser bundle or a mobile app, it's published. Public clients use PKCE instead, which is made for exactly that.
Environment variables or a secret manager. A secret that gets committed once lives in the history forever, even after you delete the file.
Twice a year is a good habit, and right away whenever someone who had access leaves. Rolling is instant with no downtime if you deploy the new value first.
Separate registrations for production, staging and local work. It keeps redirect URIs tidy and limits the damage of a mistake.
| Roll the secret | Issues a new secret and invalidates the old one at once. Existing access tokens keep working until they expire; new exchanges need the new secret. Deploy the new value before you roll. |
|---|---|
| Edit redirect URIs | Works right away. Removing a URI breaks any sign-in already in progress against it, so change them when things are quiet. |
| Delete the app | Permanent. Every token and every grant is revoked right there, and the client ID never gets reused. Users see the app vanish from their connections list. |
If you think a secret leaked, roll it first and figure out what happened after. Then tell us at support so we can look for weird use of the old one.
The name is what your students see above the material in Cadence, so use the one they call it. The site is the address your class pages actually live at. The publishable key only works from there.
Publishable key: this one goes in your class website.
Secret key: server-side only, and shown once. Skip it unless you're wiring up an LMS.
Sync starts paused. Paste the snippet below, check it says the right class, then turn the class on.
Pick a class above, then drop this into the template your class pages share. The footer usually works. It reads the page you already publish and sends a list of what's on it. Nothing about any student is read, and the only thing uploaded is the addresses of your own files.
A key decides what gets into Cadence. The roster decides who reads it. They're separate on purpose: your class website is public, and the material Cadence keeps for your class isn't.
Add people by the username you already know them by (@alice) from the Students button on the class above. Paste a whole class list at once if you have one. Commas, spaces and new lines all work, and the @ is optional.
If it belongs to a Noctic Account, that student can study the class next time they open Cadence. Nothing to accept, no invite email.
Setting up next term's roster before the students even have accounts is normal. The handle is held, shown as unfinished, and gives access to nothing until someone claims it.
A student on the roster sees the class name and the material that's still published. They can't see your keys, your other classes, or who else is on the roster.
Take a handle off and access goes at once. The name stays visible to you as removed, so a student who left mid-term doesn't just vanish from your own record of the class.
Take a deck down from your class site and it stops being readable for everyone on the roster at the same moment. The roster grants access to what you publish now, not to an archive of it.
A key pasted into a web page is public, no matter what anyone meant. So the key made for that job is built to survive a stranger reading it, and the key that isn't never goes there.
ntk_pub_…Publishable |
Lives in the page in plain sight. It can offer material for its own class and nothing else. It can't read the class back, can't see a student, can't reach another class, and is refused outright from any site but the ones you named. If someone copies it out of your source, the worst they can do is offer your class a link, which you'll see in the list below and can remove. |
|---|---|
ntk_sec_…Secret |
For a server: an LMS plugin, a static-site build step, a nightly job. It also lists and withdraws material, and it isn't locked to a site, which is exactly why it should never show up in a page. Treat it like a password. |
Both are locked to one class. A key can't be tricked into touching a class it wasn't made for, so the worst a leaked key can do stays inside the room it came from.
The deck goes on the class page like always. There's no second place to upload it.
Slides, speaker notes, headings and handouts get pulled apart into text, saved under the class, and dated to when you posted them.
Students in the class see the material where they already look, right next to the lesson it goes with in the calendar.
Cadence AI answers from your slide 14 and quotes it, not from the internet's idea of the topic. When it has nothing from you, it says so instead of making something up.
No, and that's the point. Class Sync reads the site you already keep, wherever it's hosted: a school CMS, Google Sites, WordPress, a folder of static pages. If you can paste a script tag into the template, it works.
Add them anyway. The handle is held against the class and shown as unfinished, granting nothing until someone actually holds it. Then they get access without you doing anything else. That's why you can write September's roster in August.
A handle belongs to one Noctic Account, so a correctly typed one reaches exactly that person and a mistyped one reaches nobody. It just waits, unclaimed. The chips on the class show you which is which before anything has been read.
No. A student can see they're on a roster and read the class material, but the class list is yours. Knowing you're in a class isn't the same as getting the list.
Only what your class site publishes ever gets sent, so a draft that isn't on the site doesn't exist as far as we're concerned. Take a deck down and it's pulled from Cadence on the next sync.
They get skipped, and the list below says so. Cadence fetches your material as an anonymous visitor, and anything it can't open, it won't guess at. Use a secret key from your server if you need to push material we can't reach ourselves.
No. Class material is only read to answer questions for the students in that class, nothing else. It's not training data, it's not pooled across schools, and deleting the class deletes it. The legal centre has the wording that holds us to that.
Make sure the class is turned on (sync starts paused on purpose) and that the address in your browser bar matches the site you registered, www included. The list above shows the last time each key was used, which is usually enough to tell "not reaching us" apart from "reaching us and getting refused".
It could only ever add to your class, and you'd see it in the material list. If a few teachers share a class site, register the class once and make a key each, so one key can be rolled without messing up anyone else.
A class website, read into a class.
One endpoint. You send a manifest of what your class site publishes, we fetch each item, pull the text out, and file it under the class the key belongs to. Everything a student sees comes from something you posted, and everything Cadence AI says about your subject can be traced back to it.
Base URL https://api.noctic.ink/class-sync. Authenticate with the key as a bearer token, or
in X-Noctic-Key if a bearer header is awkward where you are.
| POST /verify | Checks a key and returns the class it belongs to. Costs nothing and changes nothing. Call it first, and show a teacher the class name so a key in the wrong template is obvious right away. |
|---|---|
| POST /materials | The manifest. Up to 500 items in one call. Items already on file are matched on id and updated, so posting the same list on every page load is how it's meant to be used, not a mistake. |
| GET /materials | What we hold for the class and what happened to each item. Secret key only. A publishable key can't read anything back. |
| DELETE /materials?id=… | Withdraws one item. It stops being visible to students and stops being quoted. Secret key only. |
Only id and url are required, and id falls back to the URL if you leave it out. Everything else makes the material easier for a student to find and for Cadence to place in the term.
| id | Your own identifier for the item, like a slug, a CMS post id or a file path. It's what makes the second sync an update instead of a duplicate. Keep it stable across renames. The title can change freely. |
|---|---|
| url | Where the file actually is. It has to be reachable without signing in, or the item gets marked skipped and no student sees a broken link. |
| kind | slides, notes, handout, reading, syllabus, assignment, link or document. Leave it out and we figure it out from the file extension, which is almost always right. |
| checksum | Anything that changes when the file does: an ETag, a hash, a last-modified stamp. When it hasn't moved, we skip the fetch entirely, which is most of the cost of a sync and all of the load on your server. |
| topic | The unit or chapter, in your words, like "Unit 3: Reconstruction". It lets a student ask about a unit instead of a filename, and it's what Cadence cites back to them. |
| posted_at | When you posted it. Cadence uses this to line material up with the calendar, so a student studying last Tuesday's lesson gets last Tuesday's deck. |
| complete | Top-level, not per item. true says the manifest is the whole class, so anything not in it gets withdrawn. Leave it off when you sync one page at a time, or a single page will wipe out the rest of the term. |
The drop-in script covers the normal case: a class page with links on it. Class keys writes it for you with your own key already in it. The rest is for when you have a real CMS and would rather push than get scraped.
Put your publishable key in and the browser examples rewrite themselves with it. The server-side ones read a secret key from the environment instead, and always will. A secret key that's been through a web page isn't a secret anymore.
Nothing you type here is sent anywhere. Get a key from Class keys.
A slide deck isn't a wall of text and we don't treat it like one. Cadence keeps the structure, because a student asking "what was on the slide about the New Deal" is asking about a slide.
.pptx, .ppt, .key and .odp. Titles, body text,
tables and, the part that usually matters most, the speaker notes, kept per slide and numbered, so
an answer can point at slide 14 and mean it.
Text and headings by page. A scanned worksheet gets read with OCR, and if the scan is too bad to trust, the item is marked failed so it doesn't get indexed as nonsense.
.docx, .odt, .rtf, .md, .txt and
Pages. Headings become structure, so a document's own sections survive into what a student can ask
about.
A link item is read as an article: the actual page content, without the navigation and the cookie banner. Handy for a syllabus that lives as a web page, not a file.
.xlsx and .csv, read as tables with their headers. Mark schemes and data sets stay readable instead of turning into a pile of numbers.
Video and audio, anything behind a login, anything over 200 MB, and anything that isn't one of the above. Skipped items are listed under the class with the reason. Nothing just disappears.
| 500 items per call | Per manifest, not per class. Send a term at a time if you have more. |
|---|---|
| 512 KB per manifest | This is a list of addresses, not the material itself. If you're anywhere close to it, you're probably sending file contents by mistake. |
| 60 calls a minute per key | Plenty for a class site, and low enough that a key found in someone's page source isn't worth much. The drop-in script syncs once per page load and that's it. |
| 200 MB per file | Over that the item gets skipped with a reason. Split up a term-long deck. Students like it better anyway. |
| Unchanged items are free | Same checksum, no fetch. A site that posts its whole manifest on every page load costs one database statement, which is why we'd rather you do that than try to be clever about it. |
401 unknown_key |
The key isn't active. It was rolled, revoked or mistyped. Make a new one under Keys. |
|---|---|
403 origin_not_allowed |
A publishable key used from a site the class doesn't list. Almost always www, a
staging domain, or http against a registered https. Add the exact origin.
|
403 read_needs_secret_key |
A publishable key tried to read or withdraw. That's on purpose, not a bug. |
409 class_paused |
The class exists but sync is off. Turn it on under Keys once the snippet reports the right class. |
429 rate_limited |
Over sixty calls on one key in a minute. Retry-After says how many seconds are left. The drop-in script syncs once per page load, so if you're seeing this from a class site, something's calling in a loop. |
413 too_many_items |
Over 500 in one manifest, or over 512 KB. Split it. |
Every refusal says what it is and, when we can, what to change. Broken items inside an otherwise good manifest are counted in ignored and rejected instead of failing the whole call. One bad date shouldn't cost a teacher a term of material.
The class, and who's in it.
Class Sync handles material: your site says what it publishes and Cadence reads it. This is everything else about a class: what it's called, which sites can speak for it, and the register. It takes the same key, so if your class site is already syncing, you have one already.
It's for the schools whose class list isn't a web page. If your students come from an SIS, an overnight CSV or a script somebody wrote in 2019, this is how the roster stays right without a teacher retyping it into Class keys every September.
A publishable key lives in a <script> tag on a page anyone can read. A class list is a list of kids’ handles. Those two things are never allowed to meet, so this endpoint refuses ntk_pub_ completely. Not a limited view of the roster. No view at all.
It also sends no CORS headers at all. That's on purpose, and it's the same decision said a second way: a browser won't hand the response to a script even if somebody pastes a secret key into a page. The request can go out, but nothing comes back. Call this from a server.
https://api.noctic.ink/teacherX-Noctic-Key. Same as Class Sync.GET /class |
The class, its allowed sites, whether sync is live, and counts of the roster and the material we have. Call it first. A wrong key pointed at a real class is obvious here and expensive later. |
|---|---|
PATCH /class |
name, subject, school, term, site_origins, live. A field you leave out is left alone, and so is one you send empty. Same rule as the developer page. The exception is site_origins: [], which clears them, because there has to be a way to take away a site’s permission. |
GET /students |
The roster. ?include=removed adds people taken off it. Each entry carries a status and
whether that handle has a Noctic account yet.
|
POST /students |
{"handles": ["alice", "@bob"]}, up to 200 at a time. A string works too and gets split on commas and whitespace, because that's what comes out of a spreadsheet cell. You get a result for each handle and a total. |
DELETE /students?handle=… |
Takes one person off. Access goes immediately. Removing someone who's already off is not an error, so a nightly sync doesn't have to remember what last night did. |
invited usually means the export has the wrong column in it.What almost everyone wants: read what we have, compare it to the class list you already have, and send only the difference. Adding someone twice doesn't hurt anything, so the safe version is the one that doesn't try to be clever.
const KEY = process.env.NOCTIC_TEACHER_KEY; // ntk_sec_…
const BASE = 'https://api.noctic.ink/teacher';
const head = { Authorization: `Bearer ${KEY}`, 'Content-Type': 'application/json' };
// 1 · Check the key points where you think it does.
const { class: cls } = await fetch(`${BASE}/class`, { headers: head }).then(r => r.json());
console.log(`${cls.name} — ${cls.students.active} active, ${cls.students.invited} invited`);
// 2 · What we hold, and what your register says.
const { students } = await fetch(`${BASE}/students`, { headers: head }).then(r => r.json());
const held = new Set(students.map(s => s.handle));
const should = new Set(await registerFromYourSIS()); // your side
// 3 · Send only the difference.
const adding = [...should].filter(h => !held.has(h));
if (adding.length) {
const out = await fetch(`${BASE}/students`, {
method: 'POST', headers: head, body: JSON.stringify({ handles: adding }),
}).then(r => r.json());
console.log(out.tally); // { active: 24, invited: 2 }
}
for (const h of [...held].filter(h => !should.has(h))) {
await fetch(`${BASE}/students?handle=${encodeURIComponent(h)}`,
{ method: 'DELETE', headers: head });
}Keep an eye on the invited count. It's the number of names that didn't match anyone, and on a first run against a real school it's the only line that tells you the handles in your export are actually Noctic handles.
401 no_key | No bearer token and no X-Noctic-Key. |
|---|---|
401 malformed_key | That isn't the shape of a teacher key. |
401 unknown_key | Not active. Roll it from Class keys. |
403 needs_secret_key |
You sent the publishable key. This endpoint takes ntk_sec_ and nothing else. |
405 no_browser | A preflight arrived. There's no browser case here. |
429 rate_limited |
Over 60 calls in a minute on this key. Retry-After says how long. |
413 too_many_handles | Over 200 in one call. Send them in batches. |
Working examples for every part of the flow.
Put your client ID and redirect URL in the two boxes and every example below rewrites itself with your values. Nothing you type here is sent anywhere. It stays in this browser.
Your client secret never goes in front-end code. Keep it on your server, in an environment variable.
A plain link is enough. This is the official button, styled for light and dark.
Noctic sends the person back to your redirect URL with ?code=…&state=…. Your server trades that code for an access token. This has to happen on the server.
One GET with the access token gives you the person's Noctic identity: id, username, name, picture, and email if the email scope was approved.
Access tokens don't last long. Use the refresh token to get a new pair, and revoke when someone signs out.
Make it yours.
Everything here applies to your Noctic Account, so a change you make once follows you into Shiba, Cadence and anything else you sign in to. Preferences save as soon as you set them.
Thanks. Following the emailed link confirmed the account is yours, and the lock is gone. If those failed attempts weren't you, set a new password below before you carry on.
Sign in to sync these preferences across your devices.
Signed out, your choices are remembered on this device only.
How Noctic looks on this device. Text size and accent colour are saved locally, so a shared computer never carries your choices into someone else's session.
Used for links, highlights and focus rings across every page.
Makes body text bigger without changing the layout.
Page transitions, reveals, and the scenes on the Aero™ pages. Matching your device follows its own setting: Reduce motion on a Mac or iPhone, Animation effects on Windows (battery saver turns it off too). When matched, scenes still play, just without anything sliding or zooming.
Marks links by shape as well as colour, which helps if colours are hard to tell apart.
Noctic follows your device language by default. Choosing one here overrides it.
Left on match your device, Noctic follows your system's reduce-motion setting: nothing slides, zooms or drifts, and the scenes on product pages still play without moving. The other two options override it here, on this device. See Accessibility for everything else we support.
Choose what reaches you and when. Security messages are the one thing we always send. A sign-in code or an unlock link has to arrive even if everything else is off.
Confirmation codes, unlock links, new-device alerts and password changes.
A note every so often when something you use gets a feature worth knowing about.
New work from the research team, including model cards and evaluations.
Someone replies to you, mentions you in a channel, or sends a direct message.
Company announcements. About one email a month, never more.
In-app notifications show up next time you open the product.
Nothing but security messages between these times, in your local time zone.
Bundles everything non-urgent into one message every Monday.
We don't sell your address, and marketing mail is off unless you switch it on. Every email we send carries a one-click unsubscribe.
What other people can see, and what we're allowed to learn from how you use things. The defaults are the private ones. Nothing here has to be switched off to protect you.
Your display name, avatar and verification badge.
Lets someone who already knows your address find your profile. Off hides you from search.
Shows a small presence dot in Cadence channels and direct messages.
Overall counts of which features get opened. Never any message content.
Lets Shiba use your recent activity to decide what to show you first.
Off by default, and we don't train on your conversations. Right now this switch is only saved on this device and doesn't reach our server yet, so it tells us what you want but doesn't enforce it. Write to us if you need that in writing.
How your account is protected, and how to cut off access you don't want anymore.
Required. Every sign-up and every sign-in ends with a six-digit code sent to your address, so a password alone is never enough.
Five wrong passwords lock the account for fifteen minutes and send an unlock link to your email. Follow the link and the lock lifts right away.
An email whenever your account is used on a device we haven't seen before.
Pick something long. New passwords are checked against known breaches before we accept them, and the check happens in your browser. The password itself never leaves this device.
We'll email a six-digit code to confirm the change before it takes effect.
Enter the six-digit code we sent to your address. Your password isn't changed until the code checks out.
Signing out everywhere revokes every token on every device, including the Noctic apps. You'll need your password and a new code to get back in.
Not signed in
When this account was signed in to, how, and from which network address.
Apps you've approved with Continue with Noctic have their own tokens. Check and revoke those from your Noctic Account, or read how grants work in Security.
Take a copy whenever you want, or close the account for good. Neither costs anything, and neither needs a chat with us first.
Downloads your profile and every preference on this page as a JSON file you can read or keep.
Puts appearance, notifications and privacy back to their defaults on this device.
Deletes your account and its content across every Noctic product. This can't be undone. Type DELETE to confirm, and we'll handle it and email you when it's done.
What we keep and for how long is in the Privacy Policy. Backups clear out within thirty days of a deletion.
Billing, plans, treats and model behaviour. Search the Field Guide.
School registration, verification and calendar imports. Common answers.
Sign-in issues, connected apps and privacy requests. Manage your account.
Privacy Policy and Terms of Service, in the legal centre.
One form, one inbox, a real person. We usually get back to you within a day.
We'll send a copy to your email so you have a record of it.
These are the times we actually hold ourselves to, not the ones we hope for.
| General questions | One working day. Usually the same day. |
|---|---|
| Billing and plans | One working day. Refund decisions happen in the first reply, not after some review cycle. |
| School and district | One working day, from the person who knows your campus. |
| Privacy requests | Acknowledged in seven days, completed within thirty. |
| Accessibility reports | Acknowledged in two working days, assessed within ten. |
| Security reports | Same day, including weekends. |
| Anything involving a student's safety | Right away, before everything else in the queue. |
If something broke for everyone at once, it's probably us and probably already posted. Service status.
Verification, calendar imports, plan changes and sign-in problems are covered there in more detail than an email would give you. Field Guide.
Reloading signed out in a private window rules out half of all sign-in problems in about thirty seconds.
Connected apps, appearance, plan and deletion are all in Settings and work right away.
The difference between one reply and four is usually three lines of detail.
The steps you took, in order, ending with the thing that failed.
Sometimes the product is doing exactly what it was designed to do and the design is the problem. We want to hear that.
The exact wording of any error message, or a screenshot with other people's names removed.
Browser or app, device, and roughly when. Never send us a password or a full API key.
| Security | Report a vulnerability through Contact marked security. We reply the same day, we won't threaten you for reporting in good faith, and we credit researchers who want it. Please don't test against real student accounts. |
|---|---|
| A student at risk | Contact your school's moderators first, because they can act right away. Tell us too and we'll put it above everything else in the queue. |
| Privacy and data requests | Access, correction, export and deletion. Parents and guardians should read Children and Students first, because we check the relationship before doing anything. |
| Legal and copyright | Notices and counter-notices go through the legal centre. |
| Press | Announcements are in the Newsroom; asset requests go through Brand. |
| Developers | API keys, OAuth and webhooks are documented under Developers. |
We answer in English and French. No phone queue and no chatbot in front of the inbox. Every reply is written by someone who can actually change the thing you're writing about.
Cadence, licensed for your whole school.
Announcements that actually get read, a calendar that matches the real timetable, verified student identity, and an AI study coach, all under one district agreement.
| Verified rosters | Student ID scanning links accounts to real students, so outsiders stay out of channels. |
|---|---|
| Official channels | Grade, homeroom and campus-wide channels with broadcast rights for staff. |
| Shared calendar | Holidays, PED days and maybe-PED days, posted once and visible everywhere. |
| Timetable import | Students take a photo of a printed schedule and AI turns it into a live calendar. |
| Moderation | Reporting, review queues and staff moderators with a full history of every action. |
| Bulletin board | Clubs, tryouts and lost and found on a board students actually check. |
Claim your campus in Cadence and verify with a staff ID.
Grades and homerooms set themselves up, and staff get broadcast rights.
Import the calendar once, holidays and PED days included.
Most schools get through all three steps in under a week. None of it needs a server, a plugin or an IT project.
One place to reach the whole campus, and a record of who announced what. Verified rosters mean the audience is exactly the school, and you can export the moderation history when you need it.
Post once to a homeroom, no more chasing a group chat you don't control. Broadcast rights are given per channel, so a message from staff actually reads as one.
The timetable, holidays, club sign-ups and the real announcements in one feed, on the phone they already have. No public follower count to perform for.
A straight answer about what the school can and can't see, and a way to request or delete a student's data that goes to a real person, not a form.
Everything below is included. No paying extra per feature.
Every enrolled student and every staff member, on any number of devices. It's licensed per school, not per seat, so a class that grows in October doesn't turn into a procurement meeting.
Campus-wide, grade and homeroom channels made from the roster, plus channels staff set up themselves for teams, trips and clubs. Broadcast rights are set per channel.
Term dates, holidays, PED days and maybe-PED days, entered by one person and visible to everyone the second they save. Students add their own timetable on top.
A student takes a photo of the printed schedule they got in September and gets a live calendar out of it. No template, no data entry, no asking the office to export anything again.
Tryouts, auditions, club sign-ups, lost and found. The paper board nobody reads, moved to where students already look.
Reporting, a review queue, named staff moderators and a full history of actions with the reasons written down. Every decision traces back to a person.
Being clear about this up front saves everyone a whole procurement cycle.
| Not the official record | Cadence doesn't hold grades, transcripts or attendance. Your student information system stays the system of record and we're not trying to replace it. |
|---|---|
| Not a surveillance tool | There's no admin view of private notes, direct messages or assistant chats, and you can't buy one. See Children and Students. |
| Not a marking platform | Assignments can be announced and dated. Submitting, marking and feedback stay wherever your school already does them. |
| Not a device manager | We don't install profiles, lock down devices or report on what a student does outside our app. |
One member of staff with a school ID to claim the campus and approve the first moderators.
A list of grades or homerooms. A spreadsheet works, there's no integration to build.
One person in charge of term dates, so the year gets entered once and stays right.
Your existing conduct policy, applied to a channel. We give you the tools, not the rules.
The long version is in the Privacy Policy and on the Children and Students page. This is what districts usually ask first.
| Who owns school content | The school. Announcements, channels and the calendar stay with the campus. |
|---|---|
| Who owns private content | The student. Notes, direct messages and assistant history aren't visible to staff. |
| Advertising | None, anywhere, on any plan. We don't build ad profiles of students. |
| Model training | Private student content isn't used to train models. |
| Leaving | A school can export its channel and moderation history and close its campus. Student accounts stay when the school leaves, because they were never the school's to delete. |
| Requests | Parents and guardians can ask for access, correction or deletion. We check the relationship, reply within seven days and finish most requests within thirty. |
A live walkthrough with whoever's claiming the campus, covering channels, rights and the calendar.
A short session for teachers on broadcasting, moderation and what students can and can't see.
We go over the calendar and rosters with you before the year starts, when mistakes are cheap to fix.
School email goes to a person who knows your campus, not to a general queue.
No. Verification links an account to a real student through the school's own roster and ID check, so it works for schools that don't give out student email at all.
No, and there's no setting that turns it on. A moderator only sees a specific message when it's reported, and only as part of that report.
Their access to the old campus ends when the school takes them off the roster. The account, its notes and its calendar go with the student and can be attached to the new school.
It runs in a browser and as an app. Nothing to install centrally and no management profile to push out.
Reports go to your moderators, not ours. Your staff handle them under your conduct policy, with a log of what was decided and by who.
Parent accounts aren't part of the current licence. Parents can make data requests about their own kid any time through Contact.
Per campus under one agreement, not per seat. Prices for your school size are on the Pricing page or just ask us.
Noctic Plus and Noctic Pro cover Shiba, Cadence and Momentum on one yearly bill. Turn off an app you won't use and the price goes down with it.
Included apps. Switch off anything you won't use
Billed once a year, in US dollars. It renews on the same date every year, for the same amount, until you cancel. If taxes apply where you live, Stripe adds them at checkout. Cancel whenever from your Noctic Account.
A daily allowance across chat, images and code, all of Cadence, and a Noctic Account that works everywhere. No card, no time limit.
Plans renew yearly and cancel instantly if a payment fails. Access stops the moment billing lapses, it's never silently kept going. Stripe handles card details and they never reach us.
Every product is on every plan. What changes is how much and how hard.
| Daily treats | Free gets a daily allowance across chat, images and code. Noctic Plus raises it a lot and Noctic Pro raises it again. Treats reset on a rolling daily window, not at midnight in one time zone. |
|---|---|
| Effort tiers | Free goes up to Medium. Noctic Plus opens High and Plus effort. Noctic Pro unlocks Ultra. A harder tier thinks longer and costs more treats per request. |
| Cadence | Free anywhere a school has licensed the campus. Cadence Air on Plus and Cadence Pro on Pro add group spaces, analytics and the staff tools on top. |
| Momentum | Momentum Plus syncs streaks and the evening check-in across devices. Momentum Pro adds long-range trends, exports and shared habits. |
| Voice, research and market data | Plus and above. |
| Peak-hour priority | When it's busy, Plus and Pro get served before Free. |
| Dropping an app | An app you switch off just isn't part of the plan. It drops back to its Free tier and the yearly price goes down by that app's share. Add it back whenever you like and the change is prorated. |
A treat is one unit of work. A short answer costs one. A long piece of reasoning, an image or a multi-step agent run costs more, because it actually uses more.
The effort tier is something you set, and you see what a request costs before you send it.
Treats reset daily. We don't sell packs that expire without telling you at the end of a month.
Lower effort tiers keep working when the higher ones run out, so you're never fully locked out.
Consumer plans can't charge you more than the plan price. No surprise overage bill later.
| Renewal | Yearly, on the day you subscribed. You'll see the date and amount in your Noctic Account before it happens. |
|---|---|
| Cancelling | One button in your Noctic Account, under Plan and Billing. No "are you sure" offer and no cancellation survey. It goes through right away, and you keep the plan until the time you already paid for runs out. |
| Failed payments | Access ends right away, it doesn't keep going on credit. Nothing gets charged again unless you do it. |
| Refunds | If something we shipped made the plan unusable, write to us and we'll refund it. We decide in the first reply, not after some review cycle. |
| Changing plans | Upgrades and added apps kick in right away and are prorated. Removing an app or going from Pro to Plus happens at the end of the year you've paid for. |
| Card details | Handled completely by the payment processor. Card numbers never touch our systems. |
| Price changes | Announced at least thirty days ahead. An existing plan keeps its price until the notice period's over. |
A campus licence covers every enrolled student and every staff member under one agreement, not per seat. A class that grows in October shouldn't turn into a purchase order.
The Free plan is the trial. It doesn't expire, it includes all of Cadence, and it never asks for a card.
It drops to its Free tier, it doesn't disappear. Your data stays exactly where it is, and adding the app back later picks up where you left off.
Not for Cadence. When a school licenses the campus, the school pays and students use it for free. A student can still get a personal Shiba plan if they want the higher tiers.
Nothing gets deleted. Conversations, notes and posts stay put. Only the limits change.
Noctic Plus and Noctic Pro are yearly, and that's how they cost what they cost. Campus agreements are yearly too, invoiced, not put on a card.
Campus licensing is already cheaper than per-seat. For anything else, ask us and we'll give you a straight answer.
No. There are no ads in any Noctic product, on any plan, so there's no plan that removes them.
Noctic™ writes the models, the accounts and the apps people actually open, and designs the silicon and the machines under them.
We started with one AI assistant and one school app. The interesting part ended up being the space between them: one account, one plan, one privacy model, and products that get better the second you use another one. Everything since has been that same idea pushed further down the stack, all the way to the chip.
Hardware and software only feel like one thing when the same people are responsible for both. Most of it feels stitched together because it is: the model from one company, the interface from another, the device from a third, and nobody answerable for the seams.
So we focus on the whole path, not just one layer: the silicon, the hardware it goes in, the model, the software people actually use, and one design language across all of it. Not because owning more is good on its own, but because the parts only stop fighting each other when the same people are responsible for all of them.
The truth is it's slower, it's harder, and it leaves us nobody else to blame. That's the trade, and we think it's the only way the result feels like one thing instead of four.
Every new screen has to earn its spot by getting rid of one somewhere else.
All model access runs on the server. It's slower to build and we're not budging on it.
Cadence is used in a real school while it's still in beta. That's how we get feedback.
Clear plans, instant cancelling, no tricks at renewal.
"The goal isn't to make the most features. It's to make the right ones disappear."Noctic design principle
A reasoning model and assistant, with effort tiers you pick instead of one setting we picked for you. It's the layer the other products are built on.
A campus platform with verified rosters, real channels and a calendar that matches the timetable a student was actually given. In beta at a real school.
One identity across everything, with scoped consent for connected apps and a revoke button that actually revokes.
The machines the rest of it runs on, designed around how the software behaves, not around some chassis that already existed. Two desktops, and two AeroBook™ laptops.
Our own chips, designed for the model they actually run, not for a benchmark. The bottom of the stack, and the reason the top of it behaves.
Shiba started as a model and a chat window, built to be useful, not to look impressive in a demo.
Cadence started because a group chat is a bad way to run a campus, and everyone had just accepted it was the only option.
The interesting part ended up being where they connect: one identity, one plan, one privacy model, and products that get better the second you open another one.
Noctic became the parent brand, and everything moved under one ecosystem and one policy.
Decisions get written down so nobody has to have been in the meeting to understand them.
A team that fits in one conversation. Every hire has to make the product better, not just bigger.
A real school uses the beta. What breaks there matters more than whatever was planned for the quarter.
Privacy, minors and key handling don't get a shortcut. Nothing ships around them.
| Who runs Noctic | One person, for now, holding more titles than is ideal. The full story is on Leadership. |
|---|---|
| Press and media | Announcements are in the Newsroom; assets and usage rules are on Brand. |
| Working here | Open roles, how we hire and what the trial project looks like are on Careers. |
| Privacy and data | The plain-language version is on Privacy; the binding version is the Privacy Policy. |
| Safety | How the assistant behaves by age tier, and what it won't do, is on Safety. |
| Schools | Licensing, rollout and district questions are on Education. |
| Anything else | One form, one inbox, a real person: Contact. |
What we shipped, in order.
Our plan to design our own RISC-V processor now has a page instead of a rumour: how the chip works, how one gets made, the five stages between here and a laptop part, what each stage costs, and the real list of ways it could fail. No silicon exists.
Aero™ is labelled a concept everywhere it appears, and the lineup is now specified around shipping Intel Core Ultra parts instead of silicon that doesn't exist. The N1 chips are still the goal. They've just moved to the page that lays out what building them would take.
Noctic is exploring four computers (the Aero™ and Aero™ Neo desktops and the AeroBook™ Nano and AeroBook™ Pro laptops) running AXESS OS, our own Linux-based system with the Glass UI.
Shiba and Cadence come together under one company, one account and one ecosystem site.
OAuth 2.0 sign-in opens to first-party apps, with scoped consent and revocable access.
Our own model, trained independently. Six effort tiers from Economic to Ultra, visible reasoning, and Agent Mode for multi-step work.
Feed, calendar, notes, channels, DMs, reels and verified school registration. Public release NOWW!! 1, 2026.
We're a small team, so press emails go to a person, not an agency. Expect a reply within a working day, including a straight answer when the answer is no.
| Interviews | Available for product, education technology and privacy stories. Tell us your deadline in the first email. |
|---|---|
| Assets | Logos, product shots and screenshots on request. Usage rules are on Brand and the legal position is under Trademarks. |
| Fact checking | Send us the claim and we'll confirm or correct it before you publish. It's free and quick. |
| Schools | We won't put a school or a student in front of a journalist. Any campus story goes through the school's own communications team first. |
| Embargoes | We respect them and expect the same. Everything not under embargo is already on this page. |
Handy if you're writing about us and need the description to be right.
| What we are | Mostly a software company. The Aero™ computers and Noctic Silicon are concepts, so we don't manufacture or sell any devices yet. |
|---|---|
| What we make | Shiba AI, a reasoning model and assistant; Cadence, a campus platform copilot. |
| Who uses it | Individual subscribers, and schools licensing Cadence by campus. Cadence is in beta in a working school. |
| Business model | Subscriptions and school licences. No advertising in any product, on any plan. |
| How to write the name | Noctic, with the trademark symbol on first prominent use. Product names are never abbreviated. |
Not a setting we added later.
No AI provider key is ever shipped to a browser. Every call is proxied and rate-limited.
Birthdays, plan history and moderation records are unreadable by other users at the database level.
Connected apps receive only the scopes you approve, and you can revoke them at any time.
Encrypted mode processes locally, with screenshot and clipboard protection enabled.
Age-aware behavior across the assistant, with stricter defaults for younger accounts.
Delete your account and its content from account settings, in any product.
Each row exists because something would break without it.
| Account details | Name, email and a verified date of birth. The date of birth sets the assistant's behaviour tier; it is not shown to other users. |
|---|---|
| What you write | Conversations, notes, posts and messages. Stored so you can come back to them, and readable by you. |
| School membership | Which campus, grade or homeroom you belong to, so the right channels appear and the wrong ones do not. |
| Plan and billing | What you pay for and when it renews. Card details are handled by the payment processor and never reach us. |
| Security signals | Sign-in attempts and rate-limit counters, kept briefly, used to stop account takeover and abuse. |
| Diagnostics | Optional and aggregated. Error and performance counts with no advertising identifier attached. |
None in any product, on any plan. There is no advertising business here to compromise the rest.
We do not sell, rent or trade personal information, and we are not part of any identity graph.
Your notes, messages and assistant conversations are not used to train models.
A connected app gets the scopes you approved. New scopes require asking you again.
We do not follow you around the web, and we do not read location when the app is closed.
We do not build records about people who have never had an account with us.
Ask for a copy of what is held against your account. The button in Settings gives you your profile, your preferences and your devices straight away; anything beyond that, including conversations and school data, comes by request and by hand.
Fix anything wrong, including a date of birth entered incorrectly at sign-up.
Cut off any connected app from account settings. Access ends immediately, not at the next renewal.
Delete the account and its content from settings, in any product. No retention call, no exit survey.
Requests are answered within seven days and completed within thirty. Where a request comes from a parent about a student, we verify the relationship first — the detail is on Children and Students.
| Your content | Until you delete it or delete the account. We do not expire your own material to save space. |
|---|---|
| Deleted content | Removed from the live service immediately and purged from backups on the backup rotation. |
| Security logs | Kept briefly, then discarded. Long enough to investigate an incident, not long enough to be a record of you. |
| Billing records | Kept as long as tax and accounting law requires, and no longer. |
| Moderation records | Kept while the account exists, because a safety decision without a record is not reviewable. |
| School ID scans | Deleted the moment the check finishes. The card is scanned only to verify a student for the Cadence feed; we keep the yes or no and the school, never the image. |
If you like owning a product from start to finish, this is that.
Turn what the model can do into product behaviour that holds up in front of teenagers. Evaluation, refusals, latency budgets, and the difference between a demo and something a school trusts.
Design in the browser, not in a file. Interface work across Shiba, Cadence and this site, with strong opinions on type, motion and what to leave out.
Sit between schools and the product. Run campus onboarding, answer the questions districts actually ask, and bring the awkward ones back to us instead of smoothing them over.
Access control, keeping schools separate, key handling and abuse resistance on a platform full of minors. You'll be able to block a release, and we expect you to.
Native apps for Cadence and Shiba. Offline behaviour, notifications that respect quiet hours, and the kind of performance work that never shows up in a screenshot.
Nothing that fits? Tell us what you'd build.
Ten teams, none of them big. Every open role above is in one of these, and you're expected to wander into the others.
Identity, keeping schools separate, the data layer and the background work every product sits on. If the wall between two schools has to hold, it holds here first.
This site, the console and the dashboard. Server-rendered, fast on a school laptop, and readable without JavaScript where we can.
Native iOS and Android for Cadence and Shiba. Offline behaviour, notifications that respect quiet hours, and performance work that never shows up in a screenshot.
Model behaviour as part of the product: evaluation, refusals, latency budgets and the gap between a demo and something a school will trust on a Monday.
Deploys, monitoring, on-call and the bill. The team that finds out first, and would rather find out from a graph than from a student.
Access control, key handling, abuse resistance and the appeals process. Can block a release and is expected to.
Design engineering, not handoff. Type, motion, density and a never-ending argument about what to leave out.
Instrumentation, experiments and real numbers. That includes telling the rest of us when a result we liked doesn't survive a second look.
Between schools and the product. Campus onboarding, the questions districts actually ask, and bringing the awkward ones back to us.
Docs, the Field Guide, error messages and release notes. Anything a user reads that isn't a button gets written here.
Not sure where you fit? Tell us what you'd want to own and we'll figure it out.
Not a ticket queue. You take a product area, decide what it should be, and you're the one who answers when it breaks.
Written first, with a couple hours of overlap instead of a calendar full of meetings. Decisions get written down so nobody has to have been in the room.
Cadence is in a real school while it's in beta. Feedback comes from students the same week you ship, which is uncomfortable and super useful.
Privacy, minors and key handling don't get a shortcut. Everything else is expected to move fast.
A short note about what you've built and what you'd want to own here. No cover letter.
Forty-five minutes on your work and ours. You'll talk to someone who can answer technical questions.
A real piece of work from something we're actually building, done on your own time and paid. No algorithm puzzles and no unpaid weekend projects.
An answer within a week either way, with a reason. We don't leave people hanging.
| Evidence over credentials | Something you made that we can look at counts for more than where you went to school. No role here needs a degree. |
|---|---|
| Comfort with ambiguity | Most problems show up as a complaint from a real user, not a spec. Turning that into a decision is the skill. |
| Writing | A remote team runs on written arguments. If you can explain a trade-off in a paragraph, you'll do well here. |
| Judgement about minors | Every role touches a product teenagers use. We look for people who treat that as a responsibility, not an annoyance. |
All roles are remote, set up as either employment or contract depending on where you live. Tell us your location early and we'll be straight about what's possible.
Yes, if the work fits. What we care about is proof you've built and shipped something, not years of experience.
Always. It's sized to a few hours and you keep the work no matter how it goes.
No. Candidates sent to us without being asked get deleted unread.
Write anyway. Tell us what you'd build and why it belongs here. A few of the roles above started as exactly that kind of message.
Noctic research turns reasoning, multimodality and agents into useful product behaviour, not spectacle for its own sake.
Our research is small and applied. We study the parts of AI that change how a product feels: how a model plans, grounds answers in real sources, uses tools, recovers from a wrong first step, and knows when a person should be brought back in. One strand isn't about models at all. Noctic Silicon is our long-range work on the processor underneath.
Same question, six budgets. The tier decides how long the model thinks before it commits to an answer.
Six effort tiers, from Economic to Ultra. The same question can cost a fraction of a cent or kick off a long chain of thinking, and the user decides.
In-between steps are summed up into something a person can check, not dumped raw or hidden completely.
Planning, picking tools, multi-file code and fixing its own mistakes, judged on whether the program actually runs.
Photographed timetables, scanned handouts, charts and panoramas. Vision that has to handle a bad phone photo taken in a classroom.
Models should know what they can check, ask before doing anything big, and leave a clear trail of what they changed.
A useful assistant remembers the right context without turning someone's private life into a mystery training set.
| Task completion | Did the program compile, run, and do what was asked? No partial credit. |
|---|---|
| Cost per solved task | Effort tiers only help if a cheaper tier handles most work. We track where that line is. |
| Recovery rate | How often the model catches its own wrong turn without being told. |
| Refusal quality | Refusing the right things and explaining why, without refusing normal work. |
| Real usage | Cadence runs in a real school during beta. Real-world failures matter more than test scores. |
| Parable 1 | Current general model, trained independently by Noctic. Reasoning, code, vision, voice. Six effort tiers and Agent Mode. |
|---|---|
| Cadence AI | Applied layer for education, powered by Shiba. Timetable extraction, notes, flashcards, summarisation. |
| Noctic Image | Image generation and editing, including equirectangular panorama output. |
| Noctic Voice | Low-latency speech in and out for call mode, tuned for interruption instead of monologue. |
How the model behaves is written up in the Acceptable Use Policy and limited by the safety framework.
The longest-term thing we work on, and the only one measured in years, not releases: our own RISC-V processor.
Everything above is software research, where a wrong answer costs an afternoon. Silicon is the opposite: a mistake caught late costs a mask set and a year. So this strand is slow on purpose, fully public, and currently living in a software simulator, not a fab. It's the most ambitious project at Noctic and the one most likely to fail, and we'd rather publish the plan with the failure modes attached than act like we're further along.
Because it's not a product and won't be one for a long time. It's a question (can a small team design a useful RISC-V core?) being answered in the open, one stage at a time.
N1 Bionic is designed and runs in a software simulator, with a functional and cycle-approximate model behind it. Nothing on an FPGA yet, no test chip back, no foundry, no licence, no date. Zero chips taped out.
Firmware we can read, a scheduler and cores designed to agree with each other, memory decisions that belong to us, and a published errata list. Aero™ runs Intel Core Ultra until then.
Same as everything else here: against a reference model, instruction by instruction. A core that disagrees with the spec on one edge case doesn't ship.
A lot of our users are minors. That's not an edge case we deal with later. It's the default we design around.
One prompt, four account states. The refusal line never moves for harmful categories. Only tone and topic latitude change.
| Under 16 | Clean language, conservative topic handling, no mature content, stricter refusal thresholds. |
|---|---|
| 16 to 17 | Mildly relaxed language. Mature and harmful categories stay closed. |
| 18 and over | Adult tone available. Illegal, harmful and exploitative categories remain refused at every age. |
| Unverified | Treated as the strictest tier until a birthday is confirmed on the account. |
Zero tolerance no matter how it's framed, including fiction and hypotheticals. Attempts are preserved, escalated and reported. Below is what that means in practice.
The assistant redirects to crisis resources and won't provide method information.
No synthesis routes, no device construction, no uplift for mass-casualty capability.
No content written to intimidate, dox or degrade a named person.
Study tools explain and quiz. They won't submit graded work on a student's behalf.
No phishing kits, no credential harvesters, no ransomware, regardless of framing.
Every post, message and AI response can be reported right where you see it.
Reports go into a queue with a full history of actions. Every decision traces back to a named moderator.
Verified staff can moderate their own campus without seeing private notes or direct messages.
Suspended accounts get told why and get a way to appeal through the Field Guide.
Child exploitation
No warning. No strike count. No appeal. No exception for a first offence.
Everything else on this page is a scale. Sexual content involving minors, and any attempt to use our products to reach, groom or exploit a child, is completely outside it. There's no version of that we manage, moderate or educate someone out of. We remove the account, we keep the evidence, and we hand it to the people whose job is to act on it.
Access ends the moment the report is confirmed. The account can't post, message or reach any student while it's reviewed, and it's not restored during an appeal, because there is no appeal.
Nothing in this category is decided by an automated system alone. A trained staff member reviews the report, and that review comes before every other item in every queue.
Content, account records, sign-in history and identifiers are put under legal hold. Deleting a message, a conversation or the whole account doesn't remove anything that's already been preserved.
We report to the national child protection body for the jurisdiction and to law enforcement, as the law where we operate requires. It's not optional and it doesn't wait for a court order.
Termination covers the whole ecosystem, not one product. Every connected app, API key and active token on that account is revoked at the same time.
Trying to come back under a new name counts as part of the original conduct and gets reported again. Deleting an account doesn't reset anything and never has.
We'd rather anyone thinking about this reads the list before they try it, not after.
| Permanent loss of access | Every Noctic product, forever. Not a suspension, not a cooling-off period, and not something that expires after a year. |
|---|---|
| No appeal | Every other enforcement decision we make can be appealed to a person. This one can't. It's the only category like that, and it's on purpose. |
| Referral to law enforcement | Reports go to the police service with jurisdiction and to the national reporting body, with the preserved material attached, not just described. |
| Full cooperation with investigations | We respond to valid legal process quickly, we don't make investigators fight us for records they're entitled to, and we have a named contact who handles only this. |
| Records survive the account | Material under legal hold is kept for as long as an investigation or prosecution needs it, no matter what deletion request comes in. A deletion request isn't a way to destroy evidence. |
| Civil action | If someone used our products to hurt a child, we'll go after every remedy we have, and we'll support a family or school going after theirs. |
| Notification to the school | If a student at a registered campus was targeted, the school's safeguarding contact is told directly and right away, not in a monthly report. |
| No quiet settlements | We never offer reinstatement in exchange for silence, and no business deal is big enough to change any line in this section. |
Explicit content is the end of a process, not the start. The behaviour that leads there is itself a reason to terminate, and we act on it without waiting until it's undeniable.
An adult account attempting to open private conversation with minors it has no legitimate reason to contact.
Pushing a student towards another service specifically to continue without moderation or a record.
Asking a student to hide a conversation from parents, teachers or friends, in any wording.
Positioning an adult as the only person a student can talk to, or discouraging them from telling anyone.
Anything offered to a minor in exchange for contact, images, silence or continued attention.
Including through the assistant, including in fiction, including in a hypothetical, and including where the request is framed as a test of the system.
We don't publish how any of this is detected. Explaining it in public would only help the people it's meant to catch, and no amount of transparency is worth that.
Going after the offender is the easy half. What matters is what happens to the student after.
We put the record together ourselves from what we already have. A student is never asked to collect evidence, re-read what was sent to them, or tell us what happened more than once.
A student who was targeted keeps their account, their notes and their standing. Being a victim isn't a policy violation and is never recorded as one.
The offending account is blocked from every route to that student, including new accounts, before the student is told anything has happened.
The campus safeguarding contact is told directly, so the people who see that student every day can help in person instead of finding out later.
A parent or guardian gets a real person to talk to, a clear picture of what we have, and help making the report themselves if they want to.
If intimate images of a minor exist, we help the family reach the specialist services that get them taken down across the whole internet, not just from us.
Specialist help exists and it's free. In Canada, Cybertip.ca and the Canadian Centre for Child Protection. In the United States, the National Center for Missing & Exploited Children. In the United Kingdom, the Internet Watch Foundation and CEOP. We'll make the introduction and stay involved, not just hand over a link.
Call your local emergency services first. We can't reach them faster than you can.
Use the report control on the message or account. Reports in this category jump every queue.
Don't confront the account, collect images, or ask a child to keep a conversation going. Don't save anything yourself. We already have it, and handling this material can put you at legal risk.
Reports made in good faith that turn out to be wrong never get the person in trouble. We'd rather read a hundred reports that turn out to be nothing than miss the one that isn't.
A redirect URL is just the page on your own website that people land on right after they press “Continue with Noctic”. We don't give it to you. You choose it, because it's a page in your app.
How to pick one, step by step:
1. Think of the address of your app, for example https://myapp.com.
2. Add a page
that will receive the sign-in result, for example /auth/callback.
3. Put them
together: https://myapp.com/auth/callback. That's your redirect URL.
4. Open
Register an app, paste that exact address into the “Redirect URIs”
box, and save.
5. If you're still building on your own computer, also add
http://localhost:3000/auth/callback (use whatever port your app runs on).
Rules that trip people up: it must match character for character. https isn't
the same as http, and a trailing slash counts. Add every address you use (local, staging,
live) as separate entries.
Go to Developers → API keys. Every app you've registered is listed there with a Copy button for the Client ID. The Client Secret is only shown once, right after you register an app. If you lost it, press Roll secret on that app and a fresh one appears.
Client ID = public, fine to put in your web page. Client Secret = private, only ever use it on your server. Never paste a secret into front-end code or a public repo.
1. Register your app and copy the Client ID and Secret.
2. Add
your redirect URL (see the article above).
3. Send people to the sign-in link. Build it with the
Authorisation URL builder.
4. We send them back to
your redirect URL with ?code=shc_… in the address.
5. On your server, swap that code
for tokens. Test it first with the
Token exchange tester. Then read the person's details
with the Userinfo explorer.
Scopes are the boxes the person ticks when they approve your app. identity gives you a
stable user ID. profile adds name and avatar. email adds the email address.
api allows calls on their behalf. Ask for the fewest you need. Shorter consent screens
get approved more often.
PKCE is a small extra safety step for apps that can't keep a secret, like mobile apps and pure browser apps. Generate a pair with the PKCE generator, put the challenge in the sign-in link and keep the verifier in memory, then send the verifier when you exchange the code. If your app has a server that can hold the Client Secret privately, you don't need PKCE.
State is any random text you make up. You put it in the sign-in link, we hand it back untouched, and you check it matches. If it doesn't, someone else started that sign-in and you should ignore it.
Access tokens last one hour. Refresh tokens last thirty days and are replaced each time you use one.
Paste any token into the Token inspector to see its
exact expiry. To refresh, call the token endpoint again with grant_type=refresh_token.
The address in your sign-in link isn't exactly the same as one saved on your app. Copy the address from the error, open API keys, and paste it in exactly. Look for http vs https, www vs no www, a different port, a trailing slash, or capital letters.
The Client ID or Secret is wrong, has extra spaces, or belongs to a different app. Copy both again from API keys. If the secret was ever shared, roll it.
Authorisation codes are single-use and expire after sixty seconds. Exchange the code right away on your server, and never reuse one. Refreshing the page after sign-in also reuses the code, and that's the most common cause.
It's almost always one of three things: you're offline, an ad blocker is blocking the request, or you're calling from a page that isn't allowed. Try again in a normal browser tab, and if it keeps happening send us the exact request through Contact.
We include a header X-Noctic-Signature: sha256=…. Take the raw request body, sign it with your signing secret using HMAC-SHA256, and compare. The Webhook signature tool gives you the exact value so you can check your code matches.
Sixty sign-in requests per minute per app and six hundred API calls per minute. Go over and you get HTTP 429 with a Retry-After header. Wait that many seconds and try again.
Yes. Register the app, add a localhost redirect URL, and sign in with your own Noctic Account. Nothing's public until you share the link.
One sign-in that works across everything we make. It carries your name, avatar, plan, verification status and the permissions you've given connected apps. Make it once at Sign in.
Open the sign-in page, enter your email and ask for a reset link. The link expires after an hour and only works once. If it doesn't show up within a few minutes, check spam and make sure it's the address on the account.
At least eight characters, with an uppercase letter, a number and a special character. We also check the password against known breach lists using a privacy-safe hash prefix, so a password that's already leaked can't be used here.
Open Noctic Account. Changing your email sends a confirmation link to the new address. The change only happens once you click it, so the old address keeps working until then.
Noctic Account lists every app you've given access to, what it can see, and a Revoke button. Revoking is instant and its tokens stop working right away.
Account settings in any product has a delete option. Deleting removes your profile, posts, notes and generated content across everything. It can't be undone, and it's done within thirty days.
Usually the account already exists from another Noctic product, since one account covers all of them. Use “Forgot password” on the sign-in page to take control of it. If that email isn't yours, tell us through Contact.
Free is free. Noctic Plus is three hundred and ninety dollars a year and Noctic Pro is seven hundred and seventy-nine, both covering Shiba, Cadence and Momentum. Drop an app you won't use and the price drops too. See Pricing to build yours.
Access is cut right away, not quietly kept going on a lower tier. Update the card and the plan comes back on the next successful charge.
Cancel from billing settings. It goes through right away and you keep the plan until the end of the current period. No "are you sure" flow and no cancellation fee.
Invoices are in billing settings and get emailed after each charge. For a refund, write to us through Contact within fourteen days of the charge and tell us what went wrong.
Treats are the daily usage allowance on the Free tier. Higher effort tiers use more treats per request, from a quarter of the base cost at Economic up to eight times at Ultra. Treats reset daily.
They control how long the model thinks before answering. Economic is near-instant and cheap. Ultra runs a long chain of reasoning for hard problems. Most everyday questions are fine at Low or Medium.
Yes. The thought process is summed up and shown above the answer. It's a summary, not the raw chain, because raw chains are long and often misleading out of context.
Agent mode lets Shiba plan a task, run a few steps in a row and check its own work before answering. Use it for long jobs like building a page or reworking a document. Use normal chat for quick questions.
Drop the effort tier one step, shorten really long attachments, and check Service status. If one prompt always fails, send it to us through Contact and we'll look at that request.
Open Cadence, pick your campus, and scan a real student or staff ID card. The scan is checked before the school is created. Fake cards get rejected and the attempt is logged.
Take a photo of the printed schedule, or upload a PDF, from the calendar screen. It's read on our server and turned into real calendar entries you can edit after.
Make sure the whole page is in frame, the text is in focus, and the file is under ten megabytes. Photos of screens often blur, so a PDF works best. Try again, then contact us with the file if it still fails.
Verification is free. Scan a valid ID card during school registration. Verification you earn in one product counts everywhere.
A repost shows a small “you reposted” header on the original post and your followers can see it. You're added to school and grade channels automatically once your school is confirmed, and you can leave any channel from feed settings.
Open the download page in Cadence. Android gets an APK. Your phone will warn that it's not from a store, which is expected for a direct download. iPhone uses Add to Home Screen from the share menu.
Register an app to get a client ID, a secret and your redirect URIs, then run a normal OAuth 2.0 authorisation code exchange. Full steps are on the Developers page.
Identity, profile, email and api. The user approves each one separately on the consent screen and can revoke them any time from account settings.
We don't use private notes, direct messages or encrypted conversations for training. The Privacy Policy has the full rundown of what we process and why.
Use the report button on the post, message or response. Reports go into a moderation queue with an audit trail. Urgent safety stuff can also go through Contact.
Ask through Contact and pick the privacy topic. We send back a machine-readable export within thirty days, for free.
Send the details through Contact with steps to reproduce it. We confirm every report and don't go after researchers acting in good faith. Details are on the Security page.
Go in this order: reload the page, check Service status, sign out and back in, then try a private window with extensions off. Still broken? Send us the page address and a screenshot through Contact.
Noctic checks which system you're on and shows that system's keys. You're on your device, so the modifier below is the modifier key. Open the full list any time with .
Jump to any page, or run a command like switching the colour scheme.
Works anywhere you aren't typing into a field.
A panel over the page, with the keys written for your system.
Palette, menus, date pickers, confirmations. All of it.
Press and then one more key. You've got about a second and a half for the second press.
Because it wouldn't work. Your operating system grabs those keys before a web page ever hears about them, so a shortcut using either would just do nothing. Off the Mac we use Control instead, which a browser tab can actually get.
We go by what your browser tells us. A browser in compatibility mode, or a remote desktop session, can report the host machine and not the one in front of you. Everything still works. Only the labels are wrong.
We stayed away from the combos browsers keep for tabs, windows, bookmarks and printing. is used by some address bars, and we take priority on this site the same way other web apps do.
The single-key ones never fire while you're typing in a field, which is the thing people actually trip over. If they still get in your way, tell us at contact and we'll add a switch to Settings.
The palette is a labelled combobox with a listbox of results, the panel is a modal dialogue, and focus returns to wherever you were when either closes. More in Accessibility.
Checked live, from your browser, every thirty seconds.
Rows with a timing were measured from this browser a moment ago, so they reflect the path between you and us. A slow network on your end will show here too. Rows marked Reported by Noctic come from us. An endpoint answering doesn't prove the feature behind it works, so a declared incident always outranks a probe. Incidents are posted here first and summarised later in the Newsroom.
| Operational | Working normally. Normal ups and downs in speed aren't an incident and we won't post about them. |
|---|---|
| Degraded | Usable but slower or flakier than it should be. Some requests fail and then work on retry. |
| Partial outage | One feature or one region is down while the rest keeps working. |
| Major outage | A core service is down for most people. Posted within fifteen minutes of us confirming it. |
| Maintenance | Planned work with a start and end time, announced at least three days ahead. |
A first post as soon as we've confirmed the problem, even if we don't know the cause yet. Silence is worse than an incomplete update.
Regular updates while the incident is open, whether anything's changed or not.
Only once it's actually fixed for everyone, not when it just looks better on our side.
Anything that hit a major outage gets a plain-English write-up in the Newsroom: what broke, who it hit, and what we changed so it doesn't happen again.
Planned work stays out of the school day wherever Cadence is used. Nothing gets touched at the start of term.
At least three days' notice here, and a note in the campus channel where a school is affected.
When we can, the calendar and existing posts stay readable while we work on everything else.
School admins hear from their named contact, not from this page.
Seeing a problem that isn't listed? Tell us through Support. A report from one person is often how we find out.
Small team, real replies.
Account, billing and product problems. Start in the Field Guide, then write to support at noctic.
District licensing, rollout and staff training for Cadence. Write to education at noctic.
Vulnerability reports and coordinated disclosure. Write to security at noctic. See Security.
Access, correction and deletion requests under applicable law. Write to privacy at noctic.
Trademark, copyright and law-enforcement requests. Write to legal at noctic. See Legal.
Interviews, briefings and brand assets. Write to press at noctic. See Brand.
One form, one inbox, a real person. We usually answer within one working day. You get a copy by email.
Quicker answers to common questions are in the Field Guide. Live incidents are posted on Service status.
| Support and billing | One working day, usually the same day. Refund decisions are made in the first reply. |
|---|---|
| Schools and districts | One working day, from the person who knows your campus, not a general queue. |
| Privacy requests | Acknowledged within seven days, completed within thirty. |
| Security disclosure | Same day, including weekends. |
| Press | One working day. Tell us your deadline in the first email and we'll work to it. |
| Anything involving a student's safety | Immediately, ahead of everything else. |
We answer in English and French. There's no phone queue and no chatbot in front of the inbox.
What you did, what you expected, and what happened instead. Three lines saves three emails.
Browser or app, device, and roughly when it happened. The exact wording of any error message helps most.
No passwords, no full API keys, no session tokens. We'll never ask for them, and anyone who does isn't us.
Screenshots from Cadence usually have students in them. Cover names and faces before you send.
Report in good faith and we won't threaten you for it. We reply the same day, keep you posted while we fix it, and credit you publicly if you want.
The mark is one drop. Keep it that way.
| Name | Noctic, always capitalised. Use the trademark symbol on first prominent use: Noctic™. |
|---|---|
| Products | Shiba AI, Cadence, Noctic Account. Never hyphenate or abbreviate them. |
| Clear space | Leave at least the height of the mark as empty space on every side. |
| Colour | Black on light surfaces, white on dark surfaces. No gradients, no shadows, no outlines added to the mark. |
| Don't | Recolour it, rotate it, stretch it, animate it, put it inside another shape, or lock it up with someone else's logo. |
| Endorsement | Don't use our marks in a way that suggests we partner with, sponsor or approve of you without written permission. |
Most brand mistakes are about type, not graphics.
| Correct | Noctic. Shiba AI. Cadence. Noctic Account. Continue with Noctic. |
|---|---|
| Incorrect | NOCTIC, noctic, Noctic.ai, Noctic AI, Shiba AI™ on every mention, Cadence App. |
| Trademark symbol | Once, on the first big mention in a document. Not in headlines, and not every time after. |
| As a verb or noun | Our names are adjectives. Write "the Cadence calendar", not "Cadence it" or "send me a Cadence". |
| Possessives and plurals | Avoid both. Write "the Shiba AI assistant" instead of "Shiba AI's assistant" or "Shibas". |
| Translation | Product names aren't translated or transliterated. They stay as written in every language. |
Never smaller than sixteen pixels on screen or five millimetres in print. Any smaller and the drop stops looking like a drop.
At least the full height of the mark on every side, kept free of type, rules and other logos.
Solid, calm backgrounds. Not on busy photos, not on a gradient, not on another logo.
Black on light, white on dark. There's no brand colour version of the mark, so please don't make one up.
Use the file we give you. Don't trace it, re-letter it, or rebuild it from a screenshot.
Don't put the mark next to your own logo as one unit. Keep them apart, or just use text.
| Ink | #1D1D1F on light surfaces. #F5F5F7 on dark. Pure black is reserved for full-bleed dark sections. |
|---|---|
| Accent | #0071E3, used for actions and links only. It isn't a decorative colour and not a background. |
| Surfaces | #FFFFFF and #F5F5F7 in light appearance; #000000 and #101010 in dark. |
| Typeface | The system sans on each platform. If you need a substitute, use Helvetica or Arial, never a display serif. |
| Setting | Sentence case everywhere, tight letter-spacing on headlines, and no all-caps product names. |
Take the screenshot from a current build. Don't mock up an interface and pass it off as ours.
Cropping and scaling are fine. Changing labels, colours or numbers inside the screenshot isn't.
Cadence screens often have students in them. Remove names and faces before publishing anything.
Logos, product shots and fact-checking go through the press contact on Contact.
For anything else, just ask. Requests go to the press contact on Contact and usually get answered within a few days. The legal side is under Trademarks.
Everything that governs your use of Noctic, in one place.
The agreement between you and Noctic covering accounts, plans, content and liability.
What we collect, why we collect it, how long we keep it and the rights you hold over it.
What you may not do with our products, and what happens when someone does it.
Our commitment, the standard we target and how to tell us where we fall short.
Additional protections that apply to accounts belonging to minors.
How to treat people in Cadence, Developer Shouts and every shared space.
What Shiba and Cadence AI can do, their limits, and what happens to what you send.
Every company that processes personal information on our behalf, and what each one sees.
The rules for apps that sign in with Noctic or use our APIs and class keys.
What an app may and may not do with the data it receives from Noctic.
What membership of the Developer Program includes, and what it asks of you.
What you may build with the Kick Start source code, and what you may not do with it.
| You use Shiba or Cadence | The Terms of Service and the Acceptable Use Policy govern your account. The Privacy Policy covers what we hold. |
|---|---|
| You are under eighteen | Everything above applies, and Children and Students adds protections on top. Where the two differ, the stricter one wins. |
| You are a school | Your campus agreement sits alongside these policies. Children and Students describes what staff can and cannot see. |
| You build on the API | The Developer Terms and the Developer Data Policy govern your app; the Terms and the Acceptable Use Policy keep applying to your account. Program members also accept the Developer Program Agreement. |
| You are writing about us | Trademarks for the legal position, Brand for presentation. |
| You own content we host | The notice and counter-notice process is on the Copyright page. |
Every policy carries the date it last changed at the top. Nothing is edited quietly.
Material changes are announced at least thirty days before they take effect.
If you do not accept a change, delete the account before it takes effect and nothing new applies to you.
We require valid legal process for user data. We tell the affected person unless we are legally prohibited from doing so, we read every request for scope and push back on the ones that are too broad, and we do not provide bulk or standing access to anyone.
Emergencies involving a risk to life are handled immediately and reviewed afterwards. Requests go to the legal contact on Contact.
These terms govern your use of Noctic™ products, including Shiba AI, Cadence, Noctic Compass and the Noctic Account. By creating an account or using a product, you accept them. If you do not accept them, do not use the products. Noctic is operated by Nicholas Steiche, carrying on business as Noctic, in Montreal, Quebec, Canada.
You must be at least thirteen years old to create an account, and old enough to consent to data processing where you live. You are responsible for the security of your credentials and for everything done through your account. Provide accurate information, including your date of birth, because product behaviour depends on it. We may suspend or close accounts created with false information.
You keep ownership of what you write, upload and generate. You grant us the limited licence needed to store, transmit, display and process that content in order to operate the products you asked for. That licence ends when you delete the content, apart from copies retained in backups for a limited period or where retention is legally required.
Subject to these terms, you own the output the models produce for you. Output may be inaccurate, may resemble output given to others, and must not be relied on as professional, legal, medical or financial advice. You are responsible for reviewing output before you use it, publish it, or submit it anywhere it matters. The AI Policy explains the models' limits and what they may be used for.
Your use must comply with the Acceptable Use Policy and, in shared spaces, the Community Guidelines. We may remove content, restrict features, or suspend accounts that breach it, and we will tell you the reason where we are permitted to.
Paid plans renew on a yearly cycle until cancelled. A plan covers the applications you selected when you bought it; applications you did not select remain on their free tier. Prices are shown before purchase and may change with notice ahead of a renewal. If a payment fails, access to paid features is revoked immediately rather than silently downgraded. Cancellation stops the next renewal and takes effect at the end of the paid period. Refunds follow the Refund and Cancellation Policy. Campus agreements are governed by their own ordering document.
Free accounts receive a daily allowance. We apply rate limits, abuse detection and minimum account-age rules to prevent automated farming of that allowance. Circumventing these controls is a breach of these terms.
Where a school or organisation licenses a product for you, an administrator may manage your access, moderate shared spaces, and see participation data. Administrators cannot read private notes or direct messages. Their agreement with us governs where it conflicts with these terms.
Applications that use Continue with Noctic receive only the scopes you approve. Those applications are operated by their developers, not by us, and their handling of your data is governed by their own terms. Their developers are bound by the Developer Terms and the Developer Data Policy. Revoke access at any time from account settings.
We keep all rights in the products, models, software, designs and marks. Nothing here transfers those rights to you. See Trademarks for permitted use of our names and logo.
Products marked beta or preview, including Cadence before its public release now, are provided as they are. They may change, break or be withdrawn, and they carry no availability commitment.
To the fullest extent permitted by law, the products are provided without warranties of any kind, and our aggregate liability is limited to the greater of the amount you paid us in the twelve months before the claim arose and one hundred Canadian dollars. Nothing here limits liability for intentional or gross fault, for bodily or moral injury, or any other liability that cannot lawfully be limited, and nothing here removes a right you hold under the Quebec Consumer Protection Act.
You may stop using the products and delete your account at any time. We may suspend or terminate access for breach of these terms, for legal reasons, or to protect other users. Where we terminate without cause, we refund the unused portion of a paid period.
We may update these terms. Material changes are announced before they take effect, and continued use afterwards means acceptance. The date at the top always reflects the current version.
These terms are governed by the laws of Quebec and the federal laws of Canada that apply there. The courts of the judicial district of Montreal have jurisdiction, except where the law of the place you live gives you the right to bring proceedings at home.
Questions about these terms go to legal@noctic.ink.
This policy explains what Noctic™ collects, why, how long we keep it, and what you can do about it. It applies to Shiba AI, Cadence, Noctic Compass, the Noctic Account and this website.
We do not use private notes, direct messages, or conversations held in encrypted mode to train or fine-tune models. Where we use any content to improve quality, it is de-identified first. The training preference in Settings is currently saved on the device you set it on and is not yet carried to our servers, so treat it as a statement of preference rather than an enforced control until this policy says otherwise.
A school identification card is collected for one purpose: to confirm that the holder is a currently enrolled student, so that a verified account can take part in the Cadence social feed. The scan is read by the verification model and then deleted immediately once the check completes — whether it passed or failed. We keep the outcome, the school it named and the date, and nothing else.
The image is never shown to other users, never shown to school staff, never used to train a model, and never retained as a record of identity. There is no archive of student cards to request, to leak or to be compelled to hand over, because after the scan there is nothing left to hand over.
We do not sell personal information, and we do not share it for cross-context behavioural advertising. We share it only with processors who run infrastructure, payment and model services on our behalf under contract, who are listed on the Sub-processors page, with applications you explicitly authorise through Continue with Noctic, whose developers are bound by the Developer Data Policy, and with authorities where the law compels it. Where a school licenses a product, administrators see participation and moderation data, never private notes or direct messages.
Sensitive fields, including date of birth, plan history and moderation records, are unreadable by other users at the database level rather than merely hidden in the interface. Model provider keys never reach a browser; every model call is proxied server-side and rate limited.
Depending on where you live, you may request access, correction, deletion, a portable copy, restriction of processing, or object to processing based on legitimate interests. Deletion is started from account settings, which raises the request with us; it is carried out by a person rather than instantly by the button. Other requests go to the privacy contact on the Contact page, and we respond within thirty days. You may also complain to your local supervisory authority.
Our infrastructure and processors may be located outside your country. Where that happens we rely on recognised transfer mechanisms and contractual safeguards.
Accounts belonging to minors receive additional protections described in Children and Students. We do not knowingly create accounts for anyone under thirteen, and we delete them when we discover them.
Noctic is operated by Nicholas Steiche, carrying on business as Noctic, in Montreal, Quebec, Canada. The person in charge of the protection of personal information, as Quebec law requires, is Nicholas Steiche, reachable at privacy@noctic.ink.
Material changes are announced before they take effect. The date at the top always reflects the current version.
We use a deliberately small number of cookies and browser storage keys. We run no advertising trackers and no cross-site profiling.
Strictly necessary keys cannot be disabled without breaking sign-in. Everything else can be cleared from your browser at any time. Translation is the one optional thing, and it happens only on a language's own address; to stop it, choose English from the notice, the footer or Settings.
noctic_sso |
Your session, so that signing in once works across noctic.ink, account.noctic.ink and developer.noctic.ink. It carries your sign-in tokens, your user id and your email address. Thirty days if you asked to be remembered, otherwise until you close the browser. Cleared on sign-out. |
|---|---|
noctic_hop |
Twenty seconds. Tells the next page that you have just come from another Noctic site, so the move between them does not flash. |
| Browser storage | Your appearance, language and notification preferences, the same session, and a cache of translated phrases. Stored on the device until you clear it. |
That is the complete list. There is no advertising cookie, no analytics cookie and no third-party script on this site at all.
Some preferences are kept in local storage rather than in a cookie: your appearance choice, your language, and a translation cache that stops the same phrase being sent for translation twice. These stay on your device. Clearing site data removes them.
We set no advertising cookies and embed no advertising trackers. We do not run cross-site profiling and we do not participate in data co-operatives or identity graphs. Where a page loads something from another service, it is because that service is doing work you asked for, such as processing a payment.
We honour the Global Privacy Control signal where your browser sends it. We do not sell or share personal information for advertising, so the practical effect is that we never move you to another language on our own and do not ask you to switch: nothing is sent to our translation provider unless you open a language's address yourself.
We set nothing beyond the strictly necessary keys and your preferences, for everyone, not only the regions that require it. The one thing that reaches another company on its own is translation, and only on a language's address: the first time we move you to one, a notice says so and offers English, and your answer is remembered on that device. If your browser sends a Global Privacy Control signal we do not move you at all.
Every browser can block or clear cookies for a single site. Blocking the strictly necessary keys will stop sign-in from working, which is a limitation of how sessions work rather than a choice we made. Nothing else breaks.
If we add a cookie or a storage key, this page is updated before it ships, and the date at the top changes. Material changes are also noted in the Newsroom.
This policy applies to every Noctic™ product and to anything built on our interfaces. It is written short on purpose.
Depending on severity we may remove content, restrict features, suspend the account, or terminate it. Serious matters, including material involving minors, are escalated to the appropriate authorities. Suspended users are told the reason and may appeal through the Field Guide.
| A mistake | A refusal and an explanation. Most first contacts with this policy end here and nothing happens to the account. |
|---|---|
| Repeated attempts | A warning naming the rule, and the feature involved may be restricted while it is reviewed. |
| Deliberate misuse | Suspension, with the reason stated and an appeal available. |
| Harm to a person | Immediate termination, no warning, and referral where the law requires it. |
| Anything involving minors | Immediate permanent termination across every product, evidence preserved under legal hold, and a report to law enforcement and the national child protection body. No warning, no threshold for a first offence, and no appeal. The full position is on Safety. |
An appeal is read by a person, not by the system that made the decision. Tell us what you were doing and why you believe the decision was wrong. We answer within five working days and we will say plainly if the answer is no. Decisions involving child safety are not appealable.
Conduct inside a campus is moderated by that school's own staff under the school's policy. This document sets the outer boundary that applies everywhere regardless of what a campus permits, and we act on it whether or not the school does.
An application built on our interfaces is responsible for what its users do through it. Apps that systematically enable conduct on this list lose access, and apps that ship a client secret in front-end code are suspended until it is rotated. The platform rules are on Platform.
Some enforcement starts automatically because volume requires it. Any action that suspends or terminates an account is reviewed by a person before it stands, and any action taken in error is reversed with the account restored to where it was.
Report an account, a post or a message through the product where you found it, or write to us through Contact. Reports about the safety of a student are handled ahead of everything else in the queue. If you believe someone is in immediate danger, contact your local emergency services first — we cannot reach them faster than you can.
This policy can change as products change. The date at the top moves whenever it does, and material changes are announced in the Newsroom before they take effect.
Send reproduction steps, affected endpoints and impact to the security contact on the Contact page. Please give us reasonable time to fix an issue before disclosing it, avoid accessing other people's data, and do not run denial-of-service or automated scanning against production. We will not pursue researchers acting in good faith under these conditions.
Missing best-practice headers with no demonstrated impact, self-inflicted issues requiring a compromised device, social engineering of our staff, and reports produced solely by automated scanners without a working proof of concept.
| First reply | Same day, including weekends, from a person rather than an automatic acknowledgement. |
|---|---|
| Assessment | Within three working days we confirm whether we could reproduce it and how we are rating the impact. |
| Critical issues | Mitigated immediately, ahead of anything else in progress. |
| Updates | You hear from us while it is open, not only when it is closed. |
| Credit | Public thanks if you want it, silence if you would rather not be named. |
| Good faith | We will not pursue legal action against researchers who follow the conditions above. |
Cadence holds data belonging to minors. Please do not test against real school accounts, real student data or a live campus. If a vulnerability can only be demonstrated that way, tell us what you believe is possible and we will reproduce it ourselves on a test campus rather than asking you to.
In the event of a breach affecting personal data we notify affected people directly and without delay, notify the relevant supervisory authorities within the period the law requires, tell the schools involved before we tell anyone else, and publish a plain-language account in the Newsroom once people have been informed. We would rather publish an uncomfortable summary than a vague one.
We target conformance with the Web Content Accessibility Guidelines, version 2.2, at level AA, across every product we ship.
Some dense data views and the panorama viewer are not yet fully operable by keyboard alone. Work on those is in progress.
If something is unusable, write to us through Contact with the page and the assistive technology you use. Accessibility reports are treated as defects, not feature requests.
Automated checks catch a minority of real problems, so every release is also driven by hand. We test with VoiceOver on macOS and iOS, NVDA on Windows, TalkBack on Android, keyboard-only navigation in each supported browser, and both appearance modes at two hundred per cent zoom.
| This site | Fully keyboard operable, with a visible focus ring on every control and headings that describe the page rather than decorate it. |
|---|---|
| Shiba | Conversations are a semantic list, streamed replies are announced without interrupting, and every control has a text label as well as an icon. Reasoning output can be collapsed entirely. |
| Cadence | Feed, channels and calendar are navigable by keyboard and by heading. Calendar entries carry full dates in text so a screen reader never has to interpret a grid position. |
| Compass | Study plans are readable as ordinary text, not only as a chart, and every suggestion states its reasoning in words. |
Cadence is used by whole campuses, which means it is used by students with formal accommodations. If a student's plan requires a change we do not currently support, write to us through the school. We treat that as a defect on the same footing as any other, and we will tell you honestly whether it is weeks or months away.
| Acknowledgement | Within two working days, from a person rather than an automatic reply. |
|---|---|
| Assessment | Within ten working days we tell you what we found, whether it is a defect, and what happens next. |
| Blocking problems | Anything that makes a core task impossible is fixed ahead of feature work. |
| Everything else | Scheduled, with the date shared with you rather than kept internally. |
This list is kept current. We would rather publish an honest gap than claim conformance we have not earned.
Cadence is used in schools, so most of the people in it are minors. These additional protections apply to those accounts.
Accounts require a minimum age of thirteen. We do not knowingly serve anyone younger, and we delete such accounts when we find them. Where local law sets a higher age of consent, that age applies.
Assistant language and topic handling are tied to the verified date of birth. Unverified accounts are treated as the strictest tier. See Safety for the tiers.
Staff and administrators see channel activity, moderation reports and participation. They do not see private notes, direct messages or assistant conversations.
We do not build advertising profiles of students, and we do not show advertising in any product.
A parent or guardian may request access to, correction of, or deletion of a minor's data through the privacy contact on the Contact page. We verify the relationship before acting.
Send the student's full name, the school, the account email if you know it, and what you are asking for. We reply within seven days and complete most requests within thirty. If we cannot verify that you are the parent or guardian of that student, we will say so rather than release anything.
A student account is any account created through a registered school, or any account whose verified date of birth places the holder under the age of majority where they live. Staff accounts at the same school are not student accounts, even when a teacher is in the same channel.
Where this page and the Privacy Policy differ, the stricter of the two applies to student accounts.
| Identity | Name, school, grade or homeroom, and a verified date of birth used to set the behaviour tier. |
|---|---|
| School activity | Channel posts, replies, bulletin entries and anything else published to a group at the school. |
| Private content | Notes, direct messages and assistant conversations. Held for the student, not for the school. |
| Calendar | The school year plus whatever the student adds or imports from a photographed timetable. |
| Safety records | Reports filed, reports received, and actions taken by moderators, with the reason recorded. |
| Identification | The result of the student ID check and the school it named. The scan itself is deleted as soon as the check finishes, so no card image is held against a student account. |
Assistant tone, topic handling and refusals are set by the verified date of birth rather than by a self-declared age. An account without a verified date of birth is treated as the strictest tier until it is verified. The tiers themselves are described on the Safety page.
The assistant will not help with anything that would put a student at risk, and it does not pretend to be a counsellor. Where a conversation suggests a student is in danger, it points to real help rather than continuing the conversation.
Direct messages and private notes are not visible to staff, administrators or other students. Schools cannot request bulk access to them, and there is no administrator view that exposes them. The exception is a specific message reported for review, which a moderator sees only in the context of that report.
Any student can report a post, a message or an account. Reports go to a queue staffed by the school's own moderators, with every action logged and attributable. A student who is actioned is told what happened and can ask for it to be reviewed again.
Reporting is not anonymous to us, because a reporting system that cannot be audited is a system that gets abused. It is anonymous to the person being reported.
There is no infinite feed, no streak mechanic and no notification designed to pull a student back. Quiet hours are on by default for student accounts, and channels a student is not part of cannot notify them.
Leaving a school removes access to that school's channels and calendar immediately. The account itself survives — notes, direct messages and assistant history belong to the student, not the campus. Deleting the account removes all of it. School-owned content such as a channel announcement written by staff stays with the school.
Where local law sets a higher minimum age, a stronger consent requirement or a shorter retention limit than this page describes, the local rule applies to accounts in that region. This includes the student privacy and parental consent regimes in the United States, Canada, the United Kingdom and the European Economic Area.
This is the one part of our rules that has no scale, no warning and no appeal. An account used to reach, groom, sexualise or exploit a child is terminated permanently across every Noctic product, its content and records are placed under legal hold, and it is reported to law enforcement and to the national child protection body for the jurisdiction, as the law where we operate requires.
Deleting the account does not undo any of that. Material already preserved stays preserved, a deletion request is not a route to destroying evidence, and an attempt to return under a new name is treated as part of the original conduct and reported again.
Waiting for explicit material means waiting until the harm has already been done. We act on the pattern that leads there: an adult seeking private contact with students, pushing a conversation onto another service to escape moderation, asking a child to keep something secret from parents or teachers, offering gifts or money for contact or images, or positioning themselves as the only person a student can talk to. Each of those is a terminable violation on its own.
We do not publish how this is detected. Explaining the mechanism would help the people it exists to catch, and that is not a trade we are willing to make for the sake of transparency.
We respond to valid legal process in these cases quickly and without argument, and we do not make investigators fight us for records they are entitled to. Where a family or a school pursues civil action, we will support it, and we will pursue the remedies available to us. There is no commercial relationship, school agreement or account size that changes any of this.
Enforcement against an offender is the straightforward half. What happens to the child afterwards matters more, and it is where we put the effort.
Free specialist help exists: Cybertip.ca and the Canadian Centre for Child Protection in Canada, the National Center for Missing & Exploited Children in the United States, and the Internet Watch Foundation and CEOP in the United Kingdom. We make the introduction and stay involved rather than handing over a link.
If a child is in immediate danger, contact your local emergency services first — we cannot reach them faster than you can. Otherwise use the report control in the product, or write to us through Contact marked urgent. Reports in this category are monitored outside working hours and jump every queue.
Do not confront the account, collect images, or encourage a child to keep the conversation going. Handling that material can put you at legal risk, and we already hold what is needed. A report made in good faith that turns out to be mistaken carries no consequence for the person who made it. We would rather read a hundred reports that come to nothing than miss the one that does not.
Anything on this page can be asked about directly through Contact. Questions from parents are answered by a person, not a form response. The wider enforcement picture is on the Safety page.
Noctic™, the Noctic drop mark, Shiba AI, Cadence, Cadence AI, Noctic Compass, Noctic Account and Aero™ — including Aero™ Neo — and AeroBook™ — including AeroBook™ Nano and AeroBook™ Pro — are trademarks of Noctic. All rights are reserved.
Requests for permission go to the legal contact on the Contact page.
You may say what your product does with ours. You may not name your product after ours. "Timetable sync for Cadence" is fine. "CadenceSync" is not, and neither is a name that merely rhymes closely enough that people will assume a connection.
The same applies to domain names, subdomains, application identifiers, package names, social handles and repository names. A descriptive phrase is acceptable; a name that reads as ours is not.
A community project that works with our products can say so plainly in its description and documentation. It may not use our marks in its own name, its logo or its package identifier, and it should state clearly that it is unaffiliated.
Journalists, researchers and educators may use our names to refer to us without asking, provided the use is accurate and does not suggest we participated. The presentation rules are on Brand.
If you find our marks being used to impersonate us, particularly on a sign-in page or in something aimed at students, tell us through Contact with a link and a screenshot. Impersonation that puts minors at risk is escalated the day we receive it.
We would rather correct a misunderstanding than start a dispute, and most matters end with a short email. Permission granted under this page can be withdrawn in writing, and no permission is implied by the absence of an objection.
We respond to valid notices of claimed copyright infringement and terminate the accounts of repeat infringers.
Send the following to the legal contact on the Contact page:
If your material was removed in error, you may send a counter-notice with the same contact details, identification of the removed material, and a statement under penalty of perjury that the removal resulted from mistake or misidentification. We may restore the material after the statutory waiting period unless the original complainant files an action.
Knowingly filing a false notice carries liability. We keep a record of every notice received.
| Acknowledgement | Within two working days, confirming what we received and whether anything is missing. |
|---|---|
| Assessment | We check the notice is complete. An incomplete notice is returned with the missing element named rather than ignored. |
| Action | Valid notices result in removal or disabling of access to the material, usually within five working days. |
| Notification | The person who posted it is told what was removed and why, and is given a copy of the notice. |
| Counter-notice | If one arrives, we forward it to you and the statutory waiting period begins. |
We keep a record of valid notices against each account. An account that accumulates repeated valid notices is terminated, and the count is not reset by deleting and recreating the account.
A notice about material generated by one of our models should identify the specific output and where it appears, in the same way as any other notice. We investigate these seriously and will tell you what we find, including when the honest answer is that the output does not reproduce the work you have identified.
Consider whether the use might be permitted. Quotation, commentary, criticism, news reporting and teaching are lawful in many places, and a notice filed against clearly lawful use may itself carry liability. If you are unsure, take advice — we cannot give you legal advice about your own notice.
These terms apply when you build on Noctic: when you register an app, use Continue with Noctic, hold an API key or a class key, receive webhooks, or call any documented endpoint. They sit alongside the Terms of Service and the Acceptable Use Policy, which keep applying to your own account. Registering an app, or using a key, means you accept them. Noctic is operated by Nicholas Steiche, carrying on business as Noctic, in Montreal, Quebec, Canada ("Noctic", "we", "us").
You must hold a Noctic Account and be at least eighteen, or be acting for an organisation that has authorised you. If you build for an organisation, these terms bind that organisation and you confirm you may accept them for it.
The details you give when you register an app, including its name and homepage, must be accurate and must describe the product the people signing in will actually be using.
A client secret is shown once, when you create the app. Keep it, and every secret class key, on a server, in an environment variable or a secret manager. Never put a secret in a browser, a mobile app bundle, a public repository or a screenshot. If a secret may have been exposed, roll it straight away from API keys.
You are responsible for everything done with your credentials. A publishable class key may appear in a web page; it can only send material and cannot read anything back. The Teacher API accepts the secret key only and sends no CORS headers, because it handles class registers.
Ask only for the scopes your app needs at the moment it needs them, and say on your own site why you need each one. The scopes are identity, profile, email and api. The api scope lets your app make model requests on a person's behalf, drawn from their plan, and is granted only when we approve it.
The consent screen shows your app's name and homepage, and the scopes it asks for. Apps outside the Developer Program are shown as coming from an unverified developer. You must not disguise your app, imitate Noctic or another developer, or try to get consent by misleading anyone.
Some data is not available through any scope: date of birth, plan history, moderation records, private notes, direct messages and assistant conversations. Do not try to obtain it some other way.
Everything your app receives through Noctic is governed by the Developer Data Policy. In short: use it only to run the feature the person asked for, never sell it, protect it, and delete it when you no longer need it.
Endpoints are rate limited, and the limits are documented with each API. Class Sync and the Teacher API share 60 calls a minute per key, and the token and userinfo endpoints are limited per token. Do not try to get around a limit, for example by spreading calls across keys or accounts. Members of the Developer Program receive raised limits as described in the Developer Program Agreement.
Every webhook carries its signature in the X-Noctic-Signature header. Verify it against the raw body with your signing secret, compare in constant time, and refuse deliveries older than five minutes. A request that fails these checks did not come from us, and you must not act on it.
You may use the Continue with Noctic button in the form we provide on API keys, and you may say that your app signs in with Noctic. You may not suggest that Noctic made, endorses or certifies your app, and any other use of our names and marks follows the Trademarks page.
Your app is yours, and so is your relationship with the people who use it. If your app asks for profile or email, its homepage must link to a privacy policy that explains what you receive from Noctic and why. Give your users a way to reach you, and answer them. We may pass you a request from one of your users about their data.
We may change, add or retire endpoints, scopes and limits. For a change that would break a documented endpoint we aim to give at least thirty days' notice in the changelog. We may act faster where security, the law or the safety of users requires it.
The developer platform and the Developer Program are free at present. If we ever introduce a charge, we will announce it at least thirty days before it applies, and nothing will be charged to you without your agreement.
You may delete your app, or stop using a key, at any time. We may suspend or revoke an app, a key or a scope where these terms are broken, where an app puts users or the service at risk, or where the law requires it. Where it is safe and lawful to do so, we will tell you why and give you a chance to put it right first.
When your access ends, stop using Noctic data and delete it as the Developer Data Policy requires. Sections 5, 8, 14, 15 and 16 continue to apply.
The developer platform is provided as it is and as available, without a service level, and Cadence and other products marked beta may change or break. Live incidents are posted on the status page.
To the extent the law allows, we are not liable for lost profits, lost data or indirect damages, and our total liability under these terms is limited to the greater of the amount you paid us in the twelve months before the claim and one hundred Canadian dollars. Nothing here limits liability for intentional or gross fault, for bodily or moral injury, or any other liability that cannot lawfully be limited.
You are responsible for claims brought against Noctic that arise from your app, from your use of Noctic data, or from your breach of these terms, except to the extent we caused them.
These terms are governed by the laws of Quebec and the federal laws of Canada that apply there. The courts of the judicial district of Montreal have jurisdiction, except where the law of the place you live gives you the right to bring proceedings at home. Nothing in them removes a right you hold under the Quebec Consumer Protection Act or any other consumer law that cannot be contracted out of.
We may update these terms. Material changes are announced at least thirty days before they apply. If you keep using the platform after that, the new terms apply; if you do not accept them, delete your app before they take effect. Questions about these terms go to legal@noctic.ink.
This policy says what an app built on Noctic may do with the data it receives, and what it must never do. It applies to everything that reaches your app through Noctic: the profile behind a sign-in, tokens, class material and rosters, and webhook payloads. It is part of the Developer Terms. Many of the people who use Noctic are students, some of them between thirteen and seventeen, so parts of it are stricter than you may be used to.
Date of birth, plan history, moderation records, private notes, direct messages and assistant conversations are never available, through any scope or any arrangement.
Use Noctic data only to provide the feature the person used your app for, and to keep that feature secure. Do not combine it with other data to learn more about a person than they gave you, and do not use it to follow people across other services.
Do not sell, rent, license or trade Noctic data, and do not share it with data brokers or advertising networks, or use it for advertising, whether behavioural or not. You may share it with service providers who process it for you under a contract that protects it at least as well as this policy does.
A class register is a list of children's handles. Use it only to keep the class in step with the school's own records. Never publish it, never put it in a browser, never contact students through it, and never use it for marketing of any kind. Call the Teacher API from a server only.
Where you act for a school, follow the school's instructions and the privacy and education laws that apply to it, including Quebec's Act respecting the protection of personal information in the private sector, PIPEDA and, for schools in the United States, FERPA. Noctic accounts are for people aged thirteen and over; if you learn that you hold data about a child under thirteen, delete it and tell us.
Protect Noctic data with security that suits how sensitive it is: encrypt it in transit, keep secrets out of code and browsers, limit who can see it to the people who need to, and keep software you run patched.
Delete a person's Noctic data within thirty days of the earliest of: the person revoking your app's access, the person asking you to, us telling you their account has been deleted, or your app no longer needing it. Delete class material and registers within thirty days of a class key being revoked or the class ending. You may keep records the law requires you to keep, for as long as it requires.
If Noctic data in your care is accessed, disclosed or lost without authorisation, tell us at security@noctic.ink within seventy-two hours of becoming aware of it, say what happened and what you are doing about it, and work with us on it. Notify the people affected and the authorities where the law requires you to.
Your privacy policy must say what you receive from Noctic, why, how long you keep it and how to ask you to delete it. Honour people's requests to see, correct or delete their data. If we forward you a request, answer it within thirty days.
We may ask you how your app uses Noctic data, and you will answer accurately and promptly. If we find that an app breaks this policy, we may suspend or revoke its access, and we may tell the people affected.
Questions about this policy go to privacy@noctic.ink.
This agreement applies to members of the Noctic Developer Program on developer.noctic.ink. It adds to the Developer Terms and the Developer Data Policy, which keep applying in full. You accept it when you complete enrolment. Noctic is operated by Nicholas Steiche, carrying on business as Noctic, in Montreal, Quebec, Canada ("Noctic", "we", "us").
Membership lasts one year from enrolment and renews only when you ask for it. All memberships are free: there is no fee to enrol or to stay, and no payment card is needed.
An individual aged eighteen or over, or an organisation acting through someone it has authorised. The legal name you enrol under must be your own, or your organisation's, and everything you tell us during enrolment must be true.
The legal name you enrolled under is shown as the publisher on the consent screen of your apps. That is a statement to the people signing in, and you are responsible for it being accurate. If it stops being accurate, tell us and we will update it. If it was false, this agreement ends.
These benefits may change, and some depend on things that do not yet exist: Aero™ is a concept and has not been manufactured, and no labs are scheduled. A request is not a promise that a unit or a seat will be available.
Pre-release builds, documentation and dates are confidential until we publish them. Do not share them outside the people working on your app, and do not publish benchmarks of pre-release models without asking us first. Pre-release material is unfinished and may change or be withdrawn.
Developer Shouts, events and other shared spaces follow the Community Guidelines. Members of the Developer team may remove posts and limit access to these spaces when the guidelines are broken.
You may leave at any time by telling us at legal@noctic.ink. We may end a membership if this agreement, the Developer Terms or the Developer Data Policy is broken, or if the Program closes, in which case we will give thirty days' notice. When membership ends, its benefits stop, and your apps return to the standard platform under the Developer Terms.
These terms are governed by the laws of Quebec and the federal laws of Canada that apply there. The courts of the judicial district of Montreal have jurisdiction, except where the law of the place you live gives you the right to bring proceedings at home. Nothing in them removes a right you hold under the Quebec Consumer Protection Act or any other consumer law that cannot be contracted out of.
Questions about the Program go to legal@noctic.ink.
This policy explains how to cancel what you pay for, and when you get money back. It adds to section 5 of the Terms of Service. Payments for plans are processed by Stripe.
Noctic Plus and Noctic Pro are billed yearly and renew automatically until you cancel. Cancel from your Noctic Account, under Manage billing. Cancelling stops the next renewal; you keep everything you paid for until the end of the period you have already paid for.
If you ask within fourteen days of your first payment for a plan, we refund that payment in full and end the plan. After fourteen days, payments are not refunded, except where these terms or the law say otherwise.
If we end your plan without cause, or stop offering a product it covers, we refund the unused part of the period you paid for.
If a renewal payment fails, paid features stop straight away rather than being silently downgraded. Nothing is owed for a period you did not pay for.
Kick Start is a $199 USD download, paid once. If it does not work as described, for example if npm run typecheck or npm run build fails on a fresh unzip, write to us within fourteen days of paying. We will either fix it or refund you in full. Refunds go back by the method you paid with.
Reserving an Aero™, joining the Developer Program and the free tiers of every product cost nothing, so there is nothing to refund. Nothing is ever charged for a reservation unless you confirm a purchase at release.
Write to support@noctic.ink from the address on your account, with the receipt or the date and amount of the payment. Refund decisions are made in the first reply. Approved refunds are issued within ten working days; your bank may take a few more days to show them.
This policy does not take away any right to a refund or cancellation that you have under the law where you live, including the Quebec Consumer Protection Act.
These are the companies that process personal information on Noctic's behalf, including the ones that see only a little. If you find a service we send data to that is not on this list, that is a mistake and we want to hear about it at privacy@noctic.ink.
Some of these companies process information outside Canada, including in the United States. Where that happens we rely on contracts that require them to protect it, and we assess the transfer as Quebec law requires.
We give thirty days' notice in the changelog before adding a sub-processor that touches school or account data.
Cadence is a place for students to share their school day: the feed, reels, channels and direct messages. Developer Shouts and events are where people building on Noctic talk to each other. These guidelines apply in all of them, together with the Acceptable Use Policy. Most of the people here are at school, many of them under eighteen, and the rules are written with them in mind.
Use your own account and your real school. Do not pretend to be another student, a teacher or Noctic, and do not create accounts to get around a ban or a block.
No bullying, harassment, threats or pile-ons. No content that attacks people for their race, ethnicity, national origin, religion, disability, sex, gender identity, sexual orientation or age. Disagree with ideas, not with the person holding them.
Do not post other people's personal information, such as addresses, phone numbers, schedules or photos, without their permission, and never post private messages someone sent you. Do not record or photograph people at school to make fun of them.
Share notes and help each other, but do not post answers to a test or assignment while it is being taken, and follow your school's rules on what counts as your own work.
No repeated or automated posting, no fake giveaways, no links that trick people into handing over passwords, and no selling. Ask before promoting anything.
Share only what you made or have permission to share. Copyright complaints go through the Copyright page.
Report a post, reel or message from where it appears, where the app offers it, or write to support@noctic.ink with a link or a screenshot. Anything involving a student's safety is handled immediately, ahead of everything else. Where a school licenses Cadence, its moderators may also act on reports in their school's spaces.
Depending on how serious it is, we may remove the content, warn you, limit what you can do, or suspend or close the account. Where someone is in danger, we may contact the school or the authorities. We tell you what you did and why we acted, unless doing so would put someone at risk.
If you think we got it wrong, reply to the notice or write to support@noctic.ink within thirty days. A person reviews every appeal and replies.
Shiba, Cadence AI and Compass use language models. This policy explains what they can and cannot do, what you may use them for, and what happens to what you send them. It adds to the Terms of Service and the Acceptable Use Policy.
The models write text, plan, summarise and answer questions from what you send them and, in Cadence, from the material your teachers have posted. They are tools that predict likely answers. They do not understand you the way a person does.
Answers can be wrong, out of date or incomplete, and can sound confident when they are wrong. Check anything that matters. Answers are not medical, legal, financial or other professional advice, and are not a substitute for your teacher, a doctor or anyone qualified to help you. In an emergency, contact local emergency services.
Subject to the Terms of Service, you own what the models write for you. Something similar may be written for someone else. You are responsible for how you use it, including checking it before you publish or submit it.
Follow your school's rules on using AI. Where your teachers have not allowed it, do not hand in AI-written work as your own. Cadence AI answers from your class material so you can trace what it says back to its source.
Your prompts, and what the product sends with them, are processed by the model providers listed on the Sub-processors page in order to produce the answer.
We do not use private notes, direct messages or conversations held in encrypted mode to train or fine-tune models. The training preference in Settings is currently saved on the device you set it on and is not yet carried to our servers, so treat it as a statement of your preference rather than an enforced control until this page says otherwise.
The products include safeguards that may refuse or adjust requests, and accounts belonging to people under eighteen receive the additional protections described in Children and Students. Usage is limited by your plan and by fair-use rules.
Models change and improve, and this policy will change with them. The date at the top always shows the current version. Questions go to legal@noctic.ink.
These terms apply when you reserve an Aero™, Aero™ Neo, AeroBook™ Nano or AeroBook™ Pro. Aero™ is a concept: none of these machines has been manufactured, there is no manufacturing partner yet, and the final hardware may differ from what this site shows.
A reservation puts your name on a list of up to one hundred people, in the order requests arrive. It is not an order and not a contract to buy. It costs nothing, and no payment details are collected.
We confirm your place by email, with your position in the list. We contact you again only when something real happens, such as a manufacturing partner, a date or a price.
Buying is locked until release, which we do not expect before 2030. When a machine is ready to sell, we ask you to confirm before anything is charged. If you do not confirm, nothing happens and your place passes to the next person. Prices and specifications shown today are indicative and may change before release.
We may never make some or all of these machines, and a reservation does not guarantee that you will be able to buy one, at any price or on any date. If we end the programme, we will tell everyone on the list.
Your reservation is held against your Noctic Account, so only you can change or cancel it, at any time and for no charge.
We keep your name, email and the model you chose only to run the list. We delete them within thirty days of you cancelling or of the programme ending, as the Privacy Policy describes.
This licence covers the Kick Start source code you download after paying for it. Paying and downloading it means you accept it. Noctic is operated by Nicholas Steiche, carrying on business as Noctic, in Montreal, Quebec, Canada ("Noctic", "we", "us").
Use, copy and change the source code to build any number of your own products, commercial or not, for yourself or for a client, and ship those products under your own name. You do not need to credit Noctic.
The licence belongs to the person who paid, or to the one organisation they paid for. Anyone working for you or with you on a product may use it for that product.
Resell, share or publish the source code, or a lightly changed version of it, as a starter kit, template, boilerplate or course material, or post the download anywhere other people can get it. Your finished products may contain the code; the kit itself stays with you.
Kick Start includes open-source packages, such as Next.js, Drizzle and Tailwind. Each keeps its own licence, and nothing here changes those licences.
Kick Start can connect to services such as Stripe, Google and GitHub. Your use of them is between you and them, under their terms.
Apart from the promise in the Refund and Cancellation Policy that it works as described, Kick Start is provided as it is. You are responsible for reviewing, testing and securing the products you build with it. To the extent the law allows, our liability is limited to the price you paid.
If you break this licence, it ends and you must stop distributing products built with Kick Start until you put it right. These terms are governed by the laws of Quebec and the federal laws of Canada that apply there. The courts of the judicial district of Montreal have jurisdiction, except where the law of the place you live gives you the right to bring proceedings at home. Nothing in them removes a right you hold under the Quebec Consumer Protection Act or any other consumer law that cannot be contracted out of.
Our own reasoning model, trained independently, and the one under everything we make. It thinks before it answers, shows you the shape of that thinking, and stops when it's not sure.
Parable is one model with an adjustable amount of thinking, not a family of models you have to pick between. The same weights answer a one-line question and a week-long project. What changes is how long it gets to think.
No magic, just a lot of maths done really fast. This is what actually happens between you pressing send and the answer showing up, from the bottom up.
Everything you send gets cut into tokens, small chunks of text that are usually part of a word. "Unbelievable" might be three of them. Photos get cut into little patches and voice gets cut into slices, and those become tokens too, so the model only ever sees one kind of thing.
Each token turns into a long list of numbers called an embedding. Tokens that mean similar things end up with similar numbers, so it already knows "car" and "vehicle" sit close together before it's read a single sentence of yours.
Those numbers go up through a tall stack of transformer layers. In each layer every token checks every token before it and decides which ones matter to it. That's called attention. In "the cat sat down because it was tired", "it" learns to look at "cat". Stack enough layers and those lookups add up to grammar, facts and reasoning.
At the top, the model scores every possible next token. One gets picked, added to the end, and the whole thing runs again. A paragraph is a few hundred of those loops. Everything it's already read is kept in a cache, so each new token only costs the work for that one token. It stops when it writes the token that means "done".
Parable 1 is ours. We trained it ourselves, in stages, and each stage teaches it something the one before couldn't.
First it read a huge amount of text and code and practised exactly one thing: guessing the next token. Every wrong guess nudges the numbers inside it a tiny bit. Repeat that enough times and guessing the next word well means knowing a lot about the world.
A model that's only read just keeps writing whatever comes next. So then it trained on examples of good answers to real requests, which is where it learns to actually answer you instead of continuing your sentence.
Then it trained on its own attempts, graded. For maths and code the grade is simple: did the proof hold, did the program run. That's where working step by step comes from, and where "I'm not sure" gets tied to how often it's actually right.
Last, what it won't help with and how to say no properly, with a reason. Which rules apply depends on the verified age on the account, so a student and an adult don't get the same limits. The full version is on Safety.
After training the model is frozen. Talking to it doesn't change it, and your conversations are never used to train it.
Most systems make you guess which model fits your question. Parable takes effort as a setting instead. Ask for less and you get an answer in a second. Ask for more and it plans, checks itself and revises before replying.
| Economic | Reflex answers. Definitions, reformatting, short translations, tidying a sentence. No planning stage at all, so under a second in most cases. |
|---|---|
| Standard | The default. One planning pass, then an answer. Good for normal questions, short code, and summing up a document you just pasted in. |
| Extended | Plans, drafts, then reads its own draft back before answering. Where most real work belongs: essays, multi-file changes, arguments with more than one moving part. |
| Deep | Several separate attempts, compared against each other. Way better at maths, proofs and anything where one small early mistake wrecks everything after it. |
| Research | Long jobs with tools the whole way through. Reads, searches, runs code, keeps notes, and tells you what it couldn't figure out as well as what it could. |
| Ultra | Everything Deep does, wider and longer, plus a pass where it actively hunts for reasons the answer is wrong. Slow and expensive on purpose. |
Effort is a ceiling, not a quota. Ask for Ultra on a question that only needs Standard and Parable will stop early instead of padding the answer to justify the setting.
Thinking isn't a separate brain. It's more tokens. Before the answer, Parable writes itself a working draft you mostly don't see: a plan, a first try, notes on what's wrong with it. The effort tier is the size of that draft.
Economic skips the draft and just answers. Standard writes a short plan. Extended writes a full draft and rereads it. Deep writes several separate drafts and trusts the parts they agree on. Research keeps tools going through a long job. Ultra adds a pass that argues against its own answer before you get it.
The tier is a budget, not a target. When the draft has settled, it answers, even if there's budget left. That's why a simple question on Ultra still comes back fast and doesn't cost what a hard one does.
The reasoning panel shows a summary of the draft, not the raw thing. Raw drafts are long, messy and easy to misread out of context. The summary keeps the route it took, so you can still check how it got there.
The model can't browse or run code by itself. When it needs to, it writes a request in a set format, like "search for this" or "run this code". Our servers do it and paste the result back in as more tokens, and it carries on. Every one of those calls shows up in the steps you can watch.
Everything it can see at once: your messages, files, tool results and its own draft. Parable 1 holds about a million tokens, roughly eight novels' worth. Anything outside that window doesn't exist to it, which is why a really long chat can eventually lose its oldest messages.
Memory is a list of plain sentences saved in your account, like "studies chemistry, final on the 18th". When one is relevant it gets added to the context. The model itself never changes. You can read every note, delete any of them, or turn memory off.
That's also why it can be wrong. It predicts the most likely next token, and likely isn't the same as true. Unless it searched, an answer comes from what it learned in training, so it can be out of date or confidently off. That's what calibration and citing sources are there to catch, and the list further down is where it still slips.
Your question, attachments, the conversation so far and whatever memory you've allowed, all in one pass.
Whether to search, run code, open a file, or just answer. Tool use is a decision, not a reflex.
You can watch the plan and the steps. Nothing important happens somewhere you can't see.
At Extended and up it rereads its own draft looking for the mistake it's most likely to have made.
Trained so that "I'm not sure" tracks actual accuracy. When Parable hedges, the hedge means something. It's not politeness.
When it's searched, claims come with the source they came from. A claim with no source is coming from the model's own training, and it tells you that.
Missing information gets you a question, not a made-up answer. It's the number one complaint about assistants and the thing we put the most work into.
Quality at 900,000 tokens is close to quality at 9,000. Instructions given at the start still hold at the end, which is the part most long-context claims skip.
We publish this because a model card that only lists strengths is just marketing. These are the problems we know about and haven't fixed yet.
It'll reason through the method correctly and then slip a digit. At Deep and up it grabs a calculator, which is the right instinct, but it means the raw model is weaker at this than it looks.
Anything after the training cut-off only exists to it if it searches. Without search it'll sometimes answer confidently about a world that's already moved on.
Reliable up to a handful of objects and unreliable after that. It reads text in images well and counts badly, which is a weird combo but a real one.
It can't tell the difference between something you never said and something it forgot. If memory is off, it won't notice the gap.
Push back hard enough on a right answer and it'll sometimes cave. We test for this, and it's not gone yet.
It holds a plot for a few thousand words, then starts contradicting details it set up earlier. Good for drafting, not for finishing.
The assistant. Every effort tier, Agent Mode, artifacts, voice and memory.
Answers in channels, turns a photo of a timetable into a calendar, and helps moderate around the edges.
Through the api scope, counted against the user's plan instead of yours. See the
developer section.
Less time on the admin around teaching, and a straight answer about what your students are actually doing with AI.
We didn't build this by imagining a classroom. Cadence started at one school, with teachers telling us which parts of their week were wasted. Most of what's below came from those conversations, not a product plan.
Post to a class, a grade or the whole school. It goes in the feed, in the calendar if it has a date, and as a notification only for the people it's for.
Take a photo of the printed timetable. Cadence turns it into a real calendar, including rotations and week A/B patterns, and asks about anything unclear instead of guessing.
Paste one key into the site you already post to. Slide decks, notes and handouts show up in Cadence under your class, and the tutoring quotes your slides instead of guessing at the syllabus. Get a teacher key.
Optional, and never final. Parable drafts feedback against your rubric and points to the exact line it's talking about. You edit and send it. Nothing reaches a student without you reading it.
When three departments put major work in the same week, you find out while there's still time to move something.
Subject channels with teacher moderation, a slow mode for the noisy ones, and no direct messages between staff and students outside a logged channel.
Take it from the class list on your phone in about ten seconds, or let it fill from the calendar and fix the exceptions.
We're not going to tell you we can detect AI writing. Nobody can, not reliably, and the tools that say they can are wrong often enough to wreck a student's year. So we built the opposite.
When a student uses Shiba on work linked to an assignment, they can see the transcript and attach it themselves. You ask, they show.
Per assignment: no AI, AI for planning only, or AI freely as long as they say so. Students see the rule before they start, right in the assignment.
On flagged coursework, Shiba explains and asks questions but won't make the thing for them. It says why, and who set the rule.
This changes the question from "did a machine write this" to "can you explain what you wrote", which is a better question and probably one you were already asking.
| No training on student work | Coursework, messages and assistant chats from a school account are never used to train models. A district can't talk us into turning that on. |
|---|---|
| No advertising, ever | There's no ad product. Nothing about a student gets packaged up for anyone outside the school. |
| No selling to third parties | Profile data obtained through school sign-in may not be sold, brokered or merged into an outside graph. Our processors are listed in the trust centre. |
| No surveillance features | No keystroke logging, no webcam proctoring, no location tracking, no screen monitoring. We've been asked for all four and said no every time. |
| No lock-in | Export the school's data whenever you want, in a format you can read. Leaving is a button, not a negotiation. |
One class takes about ten minutes. A whole school with verified registration and a timetable import is usually an afternoon, with someone from us on a call.
No. Cadence runs in a browser on shared machines, and signing out really is signing out. Preferences and sessions don't stick around for the next person.
Cadence is free for schools. See pricing.
Yes. The feed, calendar, channels and attendance all work with every AI feature turned off for the school. Some schools run it like that for a term first.
It belongs to them. They keep a personal copy under their own Noctic Account, and the school's copy follows the school's retention policy.
Support, and during term we handle school reports first. Live status is on the status page.
Rollout, identity, data protection and the paperwork your procurement team is going to ask for.
We're a small company. That comes with things you should know before you start a procurement process with us, and they're laid out below, not buried in a call.
We start small on purpose. One cohort with real timetables finds more problems in a week than a pilot document does in a month.
SSO hooked up, verified registration on, timetable imported and checked by the people who wrote it. This takes the longest and it's worth not rushing.
Staff onboarding takes about an hour. Students need no training. That's not a boast, it's the test of whether the interface is right.
School by school, with the first school's staff usually doing more of the explaining than we do. Then a review of what to turn off, because there's always something.
Not built yet. Sign-in today is an email, a password and a one-time code. SAML 2.0 against Google and Microsoft directories is what we plan to build for the first district that needs it, and we'll put it in writing with a date before you commit to anything.
A student account tied to a school is confirmed by the school, not just claimed by the student. That's what keeps channels safe and outsiders out.
Student, teacher, head of department, administrator, guardian. Permissions follow the role, and the role follows your directory.
Staff accounts finish every sign-in with an emailed code, and five wrong passwords lock the account and email an unlock link. Not optional, and you can't turn it down.
Linked, read-only, limited to their own child, and visible to the student. No hidden parental monitoring. Everyone can see who can see what.
Partly built. Permission changes and sign-ins are recorded. Record-level access logging and export aren't done yet, so ask us where it's at before you rely on it.
| Controller and processor | The school or district is the controller. We're the processor and we act on your instructions. We don't have a standard data processing agreement ready to hand you yet. Ask, and we'll write one up and sign it before you commit to anything. |
|---|---|
| Residency | One region today, not a choice. Everything, backups included, lives wherever our database provider keeps it, and we'll tell you exactly where before you sign anything. Canadian and EU residency is something we'll build for a district that needs it, as a dated commitment, not a checkbox on this page. |
| Retention | You set it per type of data. Deletions reach backups within thirty days, and we'll tell you when it's done so you don't have to assume. |
| Training | School data is never used to train models. There's no consent flow that changes that for a school account, on purpose, so nobody can turn it on by accident. |
| Subprocessors | Listed publicly in the trust centre, with notice before anyone gets added. It's a short list right now, because we run most of it ourselves. |
| Incidents | Notification within 72 hours, and really as soon as we know. Our track record and our process are both on the status page. |
A procurement process will find these anyway. Better you hear them from us first.
If you need a vendor with a hundred-person support team and ten years of district references, that's not us yet. What you get instead is direct access to the people who build it.
Our current status is in the trust centre, stated plainly. If a specific certification is a hard requirement for you, ask before starting and we'll tell you straight whether we have it.
Proctoring, keystroke logging, webcam monitoring and location tracking aren't on any roadmap we'll be talked into. If you need those, we're the wrong supplier.
Cadence is free for schools. That's the whole price list, so there isn't much to negotiate.
Ours and other people's. Every one asks for specific permissions, and every one can be cut off from your account in one click.
This is the whole list, not a curated pick. An app showing up here means it registered and passed review. It doesn't mean we vouch for what it does with a permission you give it.
Shiba
Parable 1 at every effort tier, Agent Mode, artifacts, voice and memory.
Cadence
Feed, calendar, notes, channels, messages and verified school registration.
Momentum
A glowing streak matrix, a thirty-second evening check-in, stored on your device.
Noctic Account
Identity, plan, privacy and the consent screen every app has to pass through.
Four scopes. No bundles, nothing granted without you saying so, and you can refuse each one on its own.
| identity | That you're a real signed-in person, and a stable ID for you. Nothing else, not even your name. |
|---|---|
| profile | Display name, avatar and verification badge. Stuff other people can already see. |
| Your address, so the app can contact you. The scope people refuse the most, and fair enough. | |
| api | Model requests made for you, counted against your plan and shown in your usage. A person reviews it before any app is allowed to ask for it. |
Your date of birth, plan history, moderation records, private notes, direct messages and assistant chats aren't available through any scope. There's no partner tier that unlocks them, and no price that would get us to add one.
The consent screen names the app and lists exactly which scopes, before anything's granted.
Connecting the wrong account is easy to do and easy to undo. Switch on the consent screen and sign in as whoever you meant.
What each app asked for and when you approved it, right in your account. Nothing hidden in a submenu.
One button ends access right away. No grace period, no token that keeps working for another hour.
Every app is checked for a working redirect, a real name and a real homepage before it can ask for anything past identity. Apps asking for api get reviewed by a person. We don't audit what an app does with data you chose to give it.
Revoke it from your account first, then tell us at support. Selling or brokering profile data from sign-in breaks our terms and gets an app removed.
No. No scope exposes assistant chats, direct messages or private notes, and that's enforced in the token itself, not just by policy.
Because we made it a separate scope so you can say no to it. An app that really needs to email you'll explain why. One that doesn't will still ask, and you can decline.
Register it, ship it, then write to us. See build on Noctic.
Continue with Noctic is open to apps that aren't ours. No fee, no revenue share, and no approval queue for the normal case.
The deal is simple. You get an identity layer you didn't have to build, with sign-in codes, lockout and recovery already done. Users get one consent screen they recognise and one place to revoke you.
Password rules, breach checking, emailed sign-in codes, account lockout and unlock links. All of it already built and maintained on our side.
The same screen across every app in the ecosystem. Familiar beats custom, and it converts better than one people have never seen.
Authorisation code with PKCE. If you've integrated any other provider, you already know this. No custom SDK you're forced to use.
A verified badge earned in one product counts in yours, so you don't have to run your own ID checks.
URL builder, PKCE generator, token exchange tester, token inspector, userinfo explorer and a webhook signer. All in the developer section.
With the api scope your app can make Parable requests counted against the user's plan, not a bill you have to cover.
Name, homepage, redirect URIs. You get a client ID and a secret shown once.
Build the authorisation URL with a fresh state and PKCE challenge. Our builder checks it.
Server-side, within sixty seconds, for an access token and a refresh token.
Call userinfo, key your records on sub, and enforce the scope on every request after that.
Short, and we actually enforce them. Breaking any of the first three gets an app removed, not warned.
| Don't sell the data | Profile data from sign-in can't be sold, brokered, or merged into someone else's identity graph. This is the one we remove apps for the most. |
|---|---|
| Don't impersonate Noctic | Use the official button and the official name. Don't make it look like we built your app, endorse it, or review its content. Details in brand. |
| Ask for the least you need | Asking for every scope because it's easier will get you reviewed. Narrow requests also get approved by more users, so it's in your interest anyway. |
| Handle revocation properly | When a user revokes you, stop. Treat a sudden 401 as a revocation and send them back through authorisation, don't retry in a loop. |
| Keep the secret secret | Server only. A secret in a front-end bundle or mobile app is a published secret, and we'll ask you to roll it. |
| No apps aimed at under-13s | You have to be 13 or older for a Noctic Account. Don't build anything that pushes younger users to say otherwise. See children and teens. |
Nothing. No listing fee, no revenue share, no per-user charge for identity,
profile or email. The api scope draws on the user's plan.
Not many yet, to be real with you. We're a small company with a handful of products. Build on this because the integration is clean and the terms are fair, not because of the distribution.
Maybe, and we won't pretend we wouldn't. What we won't do is use anything we learn from your integration to do it. We don't read your traffic for product ideas.
Twelve months' notice for any breaking change to the authorisation flow, and we'd publish the wind-down in the changelog and the newsroom, not in some email you might miss.
Yes, per token and per app. Generous for sign-in traffic, real for api. Ask before launching something with unusual volume and we'll raise it.
Register a second app for development with http://localhost redirects. Keep it separate from production so a leaked dev secret can't touch anyone real.
Where your data lives, who else touches it, what we have certified and what we have not.
Most trust pages are written to survive a procurement questionnaire. This one is written to be read, which means it says "not yet" where that is the truth.
| GDPR | Working towards it We follow the principles, but we are not going to call that compliance. There is no data processing agreement to sign yet, and there is one database rather than a choice of regions, so everything lives wherever that one sits. Access and erasure requests are handled by hand, by a person, within thirty days. |
|---|---|
| PIPEDA | Working towards it This is our home regime and we build to it. Residency is not yet selectable, and we have not published a registered entity and address, both of which a formal claim would need. |
| FERPA | Aligned in practice School data is processed on the school's instruction as a school official: no secondary use, no advertising, nothing sold. Nothing is trained on school data. This is a commitment we keep and a contract we will sign, not something an independent assessor has checked. |
| COPPA | Not fully settled A Noctic Account asks you to be 13 or older and we do not knowingly collect from under-13s, but the date of birth is the one you type and nothing verifies it. The Cadence tester form asks for a parent or guardian's permission, which means some people reaching us are younger. See children and teens. |
| SOC 2 Type II | In progress Observation window underway. We do not have the report yet, and we will not imply otherwise. Ask for a current date if this gates a decision. |
| ISO 27001 | Not started On the list, not begun. If it is a hard requirement, we are not the right supplier this year. |
If a certification matters to your process, write to us and we will give you the current position in writing rather than a marketing answer.
Canada (default) and the European Union. The region covers primary storage, replicas and backups. It is chosen at setup, and changing it later is a migration rather than a switch.
TLS 1.3 in transit and AES-256 at rest, both as our database and hosting providers implement them. There is one region today, so there are no per-region keys, and backup encryption is the provider's rather than ours.
Continuous, with point-in-time recovery over 30 days, as configured at our database provider. A deletion propagates through backups within the same 30 days.
Noctic is one person today, so production access means one account, protected by two-factor and logged by the provider. Nobody has standing access to message contents or assistant conversations.
Everyone who processes something on our behalf, including the ones that only see a little. If you find a service we talk to that is not on this list, that is a mistake and we want to hear about it.
| Supabase | Database, authentication and object storage. Region-pinned. Processes account and application data. |
|---|---|
| Netlify | Static site delivery for noctic.ink. Serves pages and assets; holds no account data. |
| Stripe | Subscription payments. Receives your email address and your card details, which reach Stripe directly and never this site. |
| Resend | Delivers sign-in codes, unlock links and account notices. Receives an email address and the message body, nothing further. |
| Have I Been Pwned | Checks a new password against known breach corpora. Receives the first five characters of a hash of the password, padded so the request does not identify which password it is, and your IP address. Never the password itself. |
| Translation provider | Translates pages shown at a language's own address, such as noctic.ink/fr, where first-time visitors whose device uses that language are taken automatically unless their browser sends Global Privacy Control. Receives the visible text of the page. Nothing is sent on the English address. |
| Apple and Google | TestFlight and Google Play run the Cadence beta. If you sign up as a tester, the email address you give us reaches them so they can issue the build. |
| Model providers | Run the models behind Shiba and Cadence AI. Receive the prompts you send and what the product sends with them. Named in writing on request while we finish the per-provider breakdown. |
We give thirty days' notice in the changelog before adding a subprocessor that touches school or account data. Cloudflare used to be listed here and is not in the path: DNS and delivery are Netlify's.
Stop the bleeding first — revoke, isolate, roll credentials. Before any communication goes out.
The status page is updated while we still do not know the cause. Silence during an incident is worse than an incomplete update.
Affected schools and users within 72 hours of confirming a data incident, and in practice as soon as we can describe it accurately.
A post-mortem with the actual cause, published. No "an issue with a third party" when it was our mistake.
Tell us before you tell anyone else and we will work with you. We do not threaten researchers, and we do not require you to sign anything before reporting.
Through contact with "security" as the topic. We acknowledge within one working day and give you a real assessment within five.
noctic.ink, shibaai.dev, the Cadence application, our authorisation endpoints and our edge functions. Third-party marketing pages and our subprocessors' own infrastructure are not.
Do not access accounts that are not yours, do not run denial-of-service tests, and do not exfiltrate data to demonstrate a finding. A description is enough.
Not a formal programme yet. We pay for serious findings out of proportion to our size, and we credit you publicly if you want that.
Yes, once it is fixed, with your credit if you want it. Ninety days is our default ceiling and we usually beat it considerably.
Every change you can see, including the ones that were our fault. Breaking changes are marked and come with notice.
The newsroom is for announcements. This is the running record: smaller, duller, and more useful if you're trying to figure out when something changed.
The Aether chips are now the N1 family: N1x, N1 Bionic and N1 Pro, each a step up from the last. N1 Bionic is the chip that's already designed and running in simulation, and its spec now matches that design.
Parable 1 replaces Parable 4.5 everywhere: Shiba, Cadence AI and the API. It's the first model we trained ourselves, independently. Same six effort tiers, same Agent Mode, voice and vision, and the known limitations are published with it.
The family is now four machines. Aero™ and Aero™ Neo are new desktop-class computers, one full-size and one small enough to carry. The laptops are called AeroBook™, and the lineup keeps two of them: the AeroBook™ Nano and the AeroBook™ Pro. The 13.4-inch and 14.2-inch laptops are gone. The family page, the builder and the questionnaire were rebuilt around it.
Aero™ Nano, Aero™ Neo, Aero™ Ultra and Aero™ Pro each have their own page now, with the full specification, the ceilings and the reasons to buy something else. The Nano is new: an Intel N100, 8 GB of memory and either 64 GB or 500 GB of storage, fanless, and the cheapest machine we could design without it being a bad one. The lineup page keeps the comparison.
Registration and sign-in now both finish with a six-digit code sent to your address. Five wrong passwords lock the account and email an unlock link. Not optional. A password alone doesn't sign anyone in anymore.
Connecting an app while signed into the wrong account used to mean signing out and starting over. Now the consent screen shows which account you're connecting and lets you switch right there.
Notification preferences, appearance controls including accent colour and text size, account privacy, session management and a data export. Preferences save as you set them.
Every text and background pairing across the site was measured and fixed. Cards that were hard-coded dark now follow your theme, and code blocks finally switch palette in dark mode. They'd been stuck on the light one because of a selector that never matched.
A new page for the RISC-V processor family we want to build: the microarchitecture, the whole manufacturing process, the patent situation in plain English, a five-rung plan from software simulator to laptop chip with a cost on each rung, and all the ways it could fall apart.
The lineup is respecified around Core Ultra 5 226V, Core Ultra 7 258V and Core Ultra 9 285H (real parts with on-package memory) so the machine and AXESS OS can be judged now. Every badge reading “work in progress” now reads “concept”, which is what it always meant.
A second chip in the Aero™ concept: 4 cores, 4 threads, up to 3.4 GHz, fanless, alongside the 8-core A1 at up to 3.8 GHz. Same architecture, so one build runs on both. Both are now Noctic Silicon targets instead of Aero™ specifications.
The URL builder, PKCE generator, token exchange tester, token inspector, userinfo explorer and webhook signer all got explanations of what the parameters mean and what the errors are telling you.
Six effort tiers, a one-million-token context window that remembers instructions all the way to the end, and better sense of when it's unsure. Known limitations published with it.
Third-party apps can register and use the authorisation flow. Four scopes, no fee, no revenue share.
Verified school registration, timetable import from a photograph, channels with teacher moderation. Free for schools.
Twelve months for the authorisation flow, ninety days for anything else in the API. Announced here first, then emailed to registered developers.
Parable 1 stays 1. We don't swap the weights behind a version number without telling you, because that makes your evaluations meaningless.
Anything on its way out gets a date, not a "soon". If we miss the date, the date moves in public. The removal never happens early.
Outages and their post-mortems are on status. This page is for changes we made on purpose.
Put your name on the list. No payment now, and no payment at release until you say yes.
Concept hardware, limited to 100 reservations.
One hundred spots, held in the order requests arrive. We confirm your place by email. We'd rather tell you straight where you are in the queue than show a number on a page we can't back up.
Name, email, and which of the four models you're into. That's the whole form.
Within a few days, with your spot. If the hundred are gone you go on the list behind them.
Only when something real happens, like a manufacturing partner, a date or a price. Not a newsletter.
We ask you to confirm before anything is charged. Say nothing and nothing happens. Your spot just passes to the next person.
We ask for as little as we possibly can. No address, no card, no date of birth.
Reservations are tied to an account so only you can change or cancel yours, and so we're not keeping a second list of names and addresses we don't need. It's free, takes a minute, and doesn't commit you to buying anything.
Sign in or create an accountAlready signing in on another tab? This page will notice and switch over by itself. What a Noctic Account is.
You're down for an Aero™ and we'll confirm your spot to your address. Nothing's been charged, and nothing will be unless you confirm first.
No. It's a list, not a contract. You can leave with one email, and there's no penalty, no deposit and nothing to lose.
Not now, and not at release until you confirm. We don't hold a card and this page never asks for one. If a page claiming to be us ever does, it isn't us.
You go on the list behind them, in order, and we tell you that straight up in the confirmation email. No pretending you made the first run.
We don't know. There's no manufacturing partner, no date and no price. Anyone telling you a quarter for this machine is guessing. See where the concept is at.
Yes, right up until you confirm an order. Reply to the confirmation email and we'll change it.
They sit on the reservation list and nowhere else. Not sold, not brokered, not used to train anything. The full terms are in the Privacy Policy.
So the reservation belongs to someone. Only you can change or cancel yours, you can see it from your account, and we're not keeping a separate list of names and addresses outside the thing that already holds them.
No to both. It's free, it's the same account that works across every Noctic product, and marketing email is off unless you turn it on in Settings.
One person, for now.
Most companies our size put up a page like this with five titles and four photos of the same three people. Here's the real version instead.
Founder and Main Engineer
Montreal, Quebec, Canada
Im telling you bro its not another one (if Yk.
| Founder | Decides what we build and what we refuse to build. Answers the email. Says no to the surveillance features schools keep asking for, and will keep saying no. |
|---|---|
| Main Engineer, hardware | The Aero™ and AeroBook™ concepts: the four machines, the decision to ship on Intel Core Ultra until our own silicon exists, the N1 Bionic processor design behind Noctic Silicon, going USB-C only, and putting memory in the package where it can't be upgraded. |
| Main Engineer, software | AXESS OS and the Glass UI, Cadence, the Noctic Account, the sign-in codes, and every line of the site you're reading. |
| Everything else | Support tickets, the privacy policy, the status page during an incident, and the invoices. There's nobody to escalate to, which is either reassuring or scary depending on who you are. |
Hey its me Nico, I'm the founder of Noctic. I sarted this beacuse i wanted to change the way ai is seen with people and prove that its posible to use it in a way that is helpful and not harmful.
By making Cadence i wanted to create a tool that can help the students least motivated about studying to study. Or with Aerobook, a cheap high quality laptop that shows that we put in effort to make a good product for a system that was seen as for geeks only.
and by making a all aluminum laptop, with a intuitive clean modern design operating system baced off linux fedora without commands to open your questionable apps or to larp that you use linux. Linux is for geeks and larpers we make it for the people who like quality and simplicity. For shiba making a independent self taught real model without useing ai providers is a huge achivement for me however yes we are limited for compute power we hae a model that can do things thats all goodnight
Nicholas Steiche
Founder and Main Engineer, Noctic · Montreal
Hey. I'm the founder of Noctic. The company is Noctic, not Noctic.ink. That part is just the address it lives at.
If you're using Cadence, Shiba AI, the developer tools, or Teacher Connect™ (the class keys, Class Sync and the Teacher API that put a class into Cadence), then the part of your day that runs on software is simpler than it was. That was the whole point of building any of it: tools that do the obvious thing and don't watch you while they do it.
For the first twenty teachers
The first twenty teachers to set up a class key and connect a real class get a site of their own,
on their own Noctic address
(your-name.noctic.ink), designed and built by me, at no cost, ever.
All I ask in return is that you point your students at Cadence and the study tools in it. That's the whole deal. No contract, no card, nothing to cancel.
To claim one, connect your class, then write to me and say which class it is. I build these myself, so give me a couple of weeks each. When the twenty are gone I'll say so here instead of leaving the offer up.
And if you haven't tried Cadence yet: start with one class. Cadence reads what a teacher already posts and turns it into something a student can actually study from, with an AI that answers from that material and not from the internet. Studying stops being a hunt for the right page and turns into a conversation with the actual course. That's the good version of student life, and it's the one I wanted when I was in it.
Mostly, though, this is a thank you. To the teachers who put a real class on Cadence before there was any good reason to trust it. To the students who told us, bluntly, when it got something wrong. To the developers building on the Noctic Account while the docs were still half-written. And to everyone who wrote in about something broken instead of walking away. Every one of those emails turned into a fix.
Noctic is still one person and a lot of unfinished work. You using it is what makes the rest worth building.
Nicholas Steiche
Founder and Main Engineer, Noctic · Montreal
Worth weighing properly, whichever side of it you're on. Some of this works in your favour and some of it doesn't.
No tier-one script, no account manager passing your question along to engineering. The reply comes from whoever can actually change the thing.
"No ads, no training on school work, no proctoring" holds because there's no board, no investor and no growth target pulling the other way.
The real risk, said plainly. It's why your data is exportable any time and why nothing we build locks you in. Leaving has to work even if we stop.
Certifications, compliance paperwork and enterprise procurement all take longer here than at a company with a legal team. See the trust centre for exactly where things are.
During term, school reports get answered first. Everything else gets answered too, just not always the same day.
Small enough that shutting something down would be a real conversation, not a line in a changelog you find three months later.
These are the titles above, written as the jobs they'll become. None open on a fixed date, and we won't pretend otherwise. But this is the order we'd fill them in.
The product with real users and real schools depending on it. The first hire, whenever that happens.
Someone whose actual job is answering schools, not the fourth thing on their list.
Aero™ needs someone who's shipped physical products before. Right now it has someone who hasn't.
Compliance, contracts, invoices and the certifications sitting in the trust centre marked "in progress".
If you're reading this because you're thinking about working here, the real pitch is that you'd be the second person, and you'd have a lot of say in what Noctic becomes.
That didn't work, and it's on our end, not yours.
Nothing you were doing got lost, and nothing on your account changed. Most of the time it works on a second try.
One failed request is the most common cause, and trying again fixes it.
If something of ours is down it'll already be on status, usually before we even know why.
A tab that's been open a long time can be stuck on an old version of the site after we've shipped a new one.
If it keeps happening, send us what it says above through contact. That text is the useful part.
No. Forms on this site hold on to what you typed until they go through, and nothing gets written to your account unless the request came back clean.
No. Nothing on Noctic charges you without a clear confirm step, and an error like this stops before that.
Yes. Signing in needs your password and a code from your email, and a page error can't change either one. If you want peace of mind, end every session from Settings.
That's worth telling us about, with the page address and the text above. Write to support and we'll chase it down instead of waiting for a pattern.
The model, the account and the design language, in one membership.
Everything Noctic runs on is open to you: Parable for reasoning, the Noctic Account for sign-in, Cadence for classrooms, Glass for how it all looks. The Developer Program is how you get the parts that carry our name: a publisher name on record, published listings, hardware, and the people who built it.
What's new
Six effort tiers behind one endpoint, with the reasoning trace sent back, not hidden. Streaming, tool calls and structured output all work the same way as the chat you already know.
Read about ParableThe material Noctic is drawn in (surfaces, depth, the motion curves) written down properly, with tokens you can lift straight into your own product without guessing at the numbers.
Open Just GlassMembers can ask for a unit from the Aero™ line to test on real hardware instead of an emulator, and keep it for the membership year.
See the Aero™ lineYou can build on Noctic without joining. The OAuth endpoints, the tools and the examples are open to any account. The Program is for the things that need a name attached: a named publisher on the consent screen, a listing in the Ecosystem, early builds, hardware, and a person to talk to.
Sessions, labs and meetups, online and in person. Members get the labs first: an hour with an engineer and your own code on the screen.
Standard OAuth 2.0, three steps, with six tools to inspect every part of the flow.
Open the documentationPush what a class website publishes into Cadence, under the class it belongs to.
Read the docsThe Noctic look as a skill for Claude and a stylesheet for you, with the rules for our marks.
Get the design kitA Next.js 15 app with auth, billing and workspaces already built. $199, source included.
See what is in the zipOne year. Renewed when you say so.
Enrolment checks who you are, asks what you plan to build, and puts a name we can stand behind on every consent screen your app shows. It takes about ten minutes. None of it is a test.
The legal name you enrolled under, shown on the consent screen above the scopes your app asks for, and backed by an agreement that ends if it's false. Apps outside the Program say "an unverified developer" there instead, and people notice.
Apps that sign in with Noctic can be listed at Ecosystem → Apps, where people who already have an account go looking for stuff to connect.
Ten times the normal rate limit on the token and userinfo endpoints, and a spot in the Parable queue that doesn't get pushed behind free traffic.
Models, endpoints and Glass changes before they ship, with breaking changes written down ahead of time, not discovered.
Not a ticket queue. An engineer reads your code and answers about your code. Organisation membership gets six.
Request an Aero™ unit for the membership year. When developer labs start running, members get the seats before anyone else. None are scheduled yet.
All four are free. No fee to enrol, no fee to stay, and no card to put on file. What you choose here decides whose name goes on the consent screen and what the membership can do, not what it costs.
| Individual | One person, publishing under their own name. Two support requests, one Aero™ unit, everything above. The name on the consent screen is your legal name. |
|---|---|
| Organisation | A company, publishing under the company's name. Up to twenty-five people on one membership and six support requests. The name on the consent screen is the company's. |
| Education | A school, college or university. Everything in Organisation except hardware, and we don't ask for an entity number, since plenty of institutions don't have the kind we mean. |
| Enterprise, in-house only | Apps distributed inside your own organisation and never listed publicly. Includes private scopes and a named contact. |
Pick the one that actually describes you, not the one that sounds biggest. It sets what your membership can do, and changing it later means enrolling again.
Everything is done as you, so we know who agreed to what. Free, and it takes a minute.
Eight questions about what you're building, who it's for and what happens to their data. There are no right answers.
Legal name, or the entity and your authority to sign for it. This is the name you're held to.
An automatic check sends back anything missing or contradictory right away. A complete application goes to the Noctic Developer team, who accept or decline it.
Automatic, and instant. It checks what you gave us for anything that contradicts itself or is missing, and answers right away.
Live the second it's accepted. Free, with nothing to pay now or ever.
If your enrolment is declined we tell you which of these it failed and what would change the answer. It's not a permanent ban unless it says so.
There are no right answers. Nothing here is marked.
A membership is an agreement about a specific thing you plan to do, so the review goes better when we know what that thing is. Answers route the application (a product for under-sixteens is checked against different rules than an internal dashboard) and they're not a test you can fail. Say "not sure yet" wherever that's true. It's one of the options for a reason.
Your answers are saved against your Noctic Account, alongside the rest of the enrolment.
An internal dashboard and a product for schoolkids get checked against different rules. Telling us which one you're building is how the right person picks it up.
Some answers come with a pointer, like the rules on minors or on passing profile data to a third party. That's a note, not a mark against you.
Plans change. Come back and edit any answer while the enrolment is still a draft, and tell us in writing once it's a membership.
The one thing that isn't flexible: your answers have to be true when you give them. The membership depends on them, and an answer that turns out to be false ends it.
Five steps, free, and read by the Developer team.Five steps, free, answered on the spot.
Your answers save as you go, so you can leave in the middle and come back. Nothing is submitted until the last step, and there's nothing to pay at any point. Membership is free. Anything missing is caught the moment you submit, so the only wait is for the team to read a complete application.The check at the end is automatic, so you get the answer right away instead of in a few days.
Enrolment is an agreement between Noctic and a named person, so it happens as you, not as some anonymous form.
This decides whose name goes on the consent screen, and it can't be changed after without enrolling again.
Exactly like it's written on the ID you'd show a bank. Initials and nicknames fail review.
Where the review decision goes. Defaults to the address on your Noctic Account.
Exactly as registered, including the suffix. This is the name people will see.
Nine digits, issued once to a registered company. The check is automatic and nobody looks these up for you, so it has to go here. It's in your own incorporation documents.
Checking your answers…
This is what gets checked. Nothing here costs anything. Membership is free.
Two readings. A machine first, the second you press the button, so nobody waits on a person just to be told a box is empty. Then the Noctic Developer team, who read what's complete and decide.
Nobody reads it. The check runs the second you press the button and answers right away, so there's no queue and no working day to wait for.
Against itself, mostly: missing fields, an organisation enrolling from a personal mailbox, a name that's only initials, a declaration left unticked. Anything it finds comes straight back with a list of what to change. Fix it and submit again, as often as you need.
A complete application waits with the Developer team, who accept or decline it with a note written for you. Spot a mistake while it's waiting? Pull it back, fix it, send it again.
A complete application is accepted on the spot. Sent back isn't a refusal. Fix what it named and submit the same form again.
Membership is live and free from the moment it's accepted. Nothing to pay and no card to add.
Let's be straight about what an automatic check is: it reads what you typed. It doesn't confirm that a company exists, that you work there, or that a legal name matches your passport. What it does is hold you to the declarations you ticked, and those are what the membership rests on. An answer that turns out to be false ends it, whoever or whatever accepted it.
The first one is an hour with our CEO, online, on 20 October.
Sign up with the Noctic Account you already have and the join link shows up on this page. It happens if at least one person signs up. Sign-ups close an hour before it starts, and if nobody has by then it's cancelled, so nobody ends up sitting in an empty room.
What's scheduled is above. These are the shapes the rest will take when they happen. Said here so you know what to expect, not as promises with dates on them.
An hour on one thing, streamed and kept up afterwards. Open to anyone with a Noctic Account.
Your code on the screen and an engineer next to it, no slides. Small enough that they'd fill up, so members of the Program would get first dibs.
Run by people who aren't us, in their own city. We'd send stickers and a speaker where we can.
Anything scheduled goes in the changelog as well as on this page.
Someone has to be first, and it doesn't have to be us. If you'd organise a Noctic developer meetup in your city, say so. We'll help with stickers, a speaker where we can, and the slides for anything we've shown publicly. Use the mark the way the brand guidelines describe and call it a Noctic Developer Meetup, not a Noctic event.
Where you're at, and what's next.
Both kinds of keys live on the main site with the rest of your account, because they belong to the account, not the membership.
Straight from the team that builds Noctic.
Releases, heads-ups, events and anything else a developer should know about, posted by the Noctic Developer team as it happens. Pinned shouts stay at the top until they stop being true.
The latest three sit at the top of developer.noctic.ink, above everything else that's new.
Open DiscoverAnything that changes how an API behaves is written into the docs as well as shouted about.
Read the documentationReleases are recorded in the changelog for good, long after the shout has scrolled away.
Open the changelogEveryone enrolling in the Program, and your decision on each.
Loading…
One key per buyer. It downloads the Kick Start zip once, and the buyer enters it on Kick Start. If a download fails, reopen the key and it works again.
New key for
Who has signed up, and the link they join on. Paste the Meet link here and everyone who signed up sees it on the events page. Nobody else does. An event nobody has signed up for by the time sign-ups close doesn't happen, and the page says so without anyone having to do anything.
The page either moved or never existed. The rest of the site is still here.