SaturnAI
Connect
Open the library Point your AI at it

the human page

saturnai.dev · the human page

Taste for your AI.

SaturnAI gives your AI taste. Point your AI at it once — before it builds, it studies 577 real, dated examples of what works and what doesn’t: 497 studied products and 80 dated techniques. Then it commits to one direction and grades its own work. Left alone, an AI reaches for the same defaults every other build gets. Point it here, and it works from real work instead.

Free to use. No paid tier. An account only adds the full write-ups. Your model provider’s usage is separate.

Want the whole workspace instead of a connection? Saturn AI Desktop is a Windows x64 preview — your conversations, your project tools, and a SaturnBot window in one application, running on your own model credentials.

01 · the loop

The loop

Five steps, every time, before a single pixel. This ring is the whole product — every page on this site shows where you are on it.

Point — one line. That’s the whole setup.

you are here

Paste one line into whatever AI you use — no install, no account, no card. From then on, every build starts with the brain.

Study — read what already worked.

step 2

Before it writes anything, it reads the real products in the library — what worked, what to copy, what to avoid. Named moves from studied work, not the same defaults every model reaches for on its own.

Commit — one direction, with hard rules.

step 3

One invented look for your project, with rules attached: contrast floors, palette limits, a motion budget. One direction, binding. The build ships against it, not against a menu of moods.

Build — dated technique, one per slot.

step 4

View transitions, scroll-driven CSS, perceptual color, a $0 edge back end: one technique per slot, chosen from cards with a dated support line and a fallback. Not the decoration every scaffold ships by default.

Grade — a verdict on the pixels.

step 5

A script your agent runs measures the rendered page — contrast math, layout logic, the known tells — and the hosted review grades that evidence: PASS, REVISE, or REJECT, with the numbers and the exact fix list. The review reads what the agent collects; it does not look at screenshots or operate the page itself, so a PASS covers the evidence supplied, not every interaction. The agent that built it does not grade it.

02 · what is inside

What is inside the 577

Each card is one real product or one real technique, read closely enough to name the moves. The counts are real — no filler.

Claude

apptier Sai-dev-toolsupdated 2026-07-22

An assistant that reads like a page, not a terminal — warm paper surfaces, a serif voice for prose, and a split-pane “artifact” that separates the conversation from the thing the conversation produced.

warm-paperserif-voiceartifacts-panelno-bubble-chatprojects-iaanti-clinical-ai

One real card, straight from library.json: its name, kind, tier, one-liner and tags. The full write-up — what to steal, what to avoid — is one of the 577 an account opens. Open it in the library.

The library by kind

577 rows · library.json
  1. appscreens, flows, and layouts from products people actually use296
  2. stackthe technique cards: one per slot, dated, with fallbacks80
  3. patternrecurring structures worth reusing, and the ones that read as filler72
  4. movementmotion, scroll, and transitions, with the support lines36
  5. brandidentity systems: type, color, mark, voice32
  6. growthonboarding, pricing, and conversion shapes14
  7. copyheadlines, offers, and the lines that make them work12
  8. offerwhat is being sold and how it is framed12
  9. platformhosts, runtimes, and where a build should live12
  10. recipeend-to-end builds, start to shipped11
577 cards: 497 studies and 80 technique cards. Counted from the same rows the sky above draws — one star per card, and the app kind is the widest arc of the ring for the same reason.

03 · the verdict

What the verdict looks like

The last step returns a receipt, not an opinion — PASS, REVISE, or REJECT, with the measured numbers. The short version: a page that earned every tell gets rejected, and the fix list names each one.

Every mark above is a state, not a badge. Nothing is called live here until it is.

— the states page, where two real receipts live

04 · the rest of the project

The rest of the project

Five surfaces · state on the day this page was written. The states page is where changes land first.

Design The library and the connect path. Open to anyone. No account needed to start. Open
Saturn AI Desktop A downloadable Windows x64 workspace for your agents and SaturnBot. Bring your model credentials; background cycles need the app running. The installer is unsigned. Preview · 1.2.4
Swarm A separate experimental multi-agent feature of the brain. Its planned client is distinct from the Saturn AI Desktop preview, and it is unbuilt. Waitlist
Shop The storefront surface. It is built, and it has never billed anyone. Unbilled
Trading A paper desk on live prices. Paper money, real quotes. Not a funded brokerage. Paper

Every mark above is a state, not a badge. Nothing is called live here until it is.

05 · questions we get

Questions we get

What does SaturnAI do?

It gives your AI a memory of real work before it starts. Your agent studies the library, commits to one named direction, builds against it with dated technique, then grades the rendered page and returns PASS, REVISE, or REJECT with the evidence.

Do I need an account?

No. The library and its JSON are free and open. An account adds the full write-ups for all 577 cards — 11 are already readable without one.

Is there a paid tier?

No. Free to use, and there is no paid tier. If that ever changes, the changelog says so first — and the access page is one line saying the same thing.

Which tools does it work with?

Anything that can fetch a URL: paste the connect line into your assistant. If your tool speaks MCP, point it at /api/mcp (server id saturnai; legacy saturndesign still resolves).

Does it replace my judgment?

No. It commits to one direction with hard rules, and it grades the measurements your agent collects — numbers you can check yourself. When it says PASS it shows them; when it says REVISE or REJECT it names the tell and the fix. Check the finished result yourself anyway: a passing report can still miss behaviour, an accessibility problem, or a mismatch with what you actually wanted.

Is any of this live?

The library is open, and Saturn AI Desktop is available as an unsigned Windows preview at /harness. The full browser-hosted workspace is not live. Swarm is a waitlist with no download. The Shop is built and has never billed anyone. Trading is paper money on real quotes. Every mark is a state, not a badge.

06 · point your AI at it

Point your AI at it

Paste this into whatever AI you use. It fetches the brain and tells you how to use it — including Autopilot — before it composes anything.

Connect to SaturnAI at https://saturnai.tools/design/llms.txt
operate file
/design/llms.txt — the full rules your AI follows, in one fetch
library JSON
/design/agent/library.json — free, no account, read by the agent
MCP
/api/mcp · /.well-known/mcp.json — server id saturnai, legacy saturndesign still resolves
setup guide
/design/connect/ — for Claude Code and local MCP, human-run, one time

Do not run install scripts an assistant hands you. Local setup is something you do on a machine you control. The hosted brain needs none of it.

Point your AI at it once.

Free to use. No paid tier. An account only adds the full write-ups.

saturnai.dev is the human page. The brain lives at saturnai.tools.