SaturnAI
Connect
saturnai.dev Connect your agent

the agent brain

saturnai.tools · the agent’s door

The agent harness that proves its work.

The host an AI assistant fetches before it builds UI, brand, growth, copy or pricing: a 577-card library of studied real work, one committed direction, and a measured verdict before anything ships.

Free. Anonymous: 60 calls an hour. A free izzy.la account opens full card DNA and 600 an hour. There is no paid tier.

Give it a brief. It answers with a contract.

Not a demo. This form posts to POST /api/design/call — the same endpoint your assistant calls — and prints what came back: one committed direction, the decoration it refuses, and the technique slots its build has to fill. The cards it drew on come forward in the sky above.

Nothing is stored. The response is the real tool output, trimmed to four fields; your assistant receives the whole object, including the DTCG tokens and the build checklist.

01 · the loop

Three tools. The direction it returns is binding.

An assistant connected to this host runs the same loop on every design task: intake before anything is built, compose for one direction, review after the build. Writer never grades writer.

intake — before anything is built

20 fields

Turns what the human actually said into a 20-field context ledger and asks about every field it cannot infer, recommended answer first. Only when all 20 are settled does it write the builder prompt — which calls compose. Nothing gets built from a one-line brief.

compose — the direction

binding

Returns one named direction with typed constraints and DTCG tokens, plus premium_stack: the technique ladder for the brief’s surface — required slots with alternatives, recommended reaches, and the priced-in decoration to refuse. It also returns announce.say, a plain-words line to relay before building. The direction is binding, not a mood board to sample from. Never clone one card.

review — after the build

fresh context

Measures the rendered page — WCAG contrast maths, layout logic per section, technique depth against the required slots, and priced-in decoration — and returns PASS, REVISE with the exact fix list, or REJECT. It runs from a fresh context, because the writer never grades its own work.

Premium is the default, not the upgrade: one technique per required slot, or the deviation is declared with a reason. Every technique card costs $0 to run.

02 · the verdict — real receipts

Every claim on this page is a receipt the engine actually issued.

After a build, a fresh context runs collect-evidence.js in the rendered page — computed colours, contrast maths, layout logic, technique probes — and posts the measurements to review. What comes back is the object below. These two were issued against the engine’s own review fixtures on the date each card names; the rubric rows, the scores and the evidence lines are verbatim.

PASS test/fixtures/premium.html graded 2026-09-15 · review via saturnai.tools/api/design/call

100% evidence coverage · 5 contrast pairs measured, lowest 6.09:1 · no slop tell fired · 10 technique slots proven, 4 declared, 0 missing

slop_avoidance
5
PASS
accessibility
5
PASS
system_coherence
5
PASS
direction_adherence
5
PASS
distinctiveness
5
PASS
technique_depth
4
PASS

No blocker. Every check that could run, ran.

REJECT test/fixtures/slop.html graded 2026-09-15 · review via saturnai.tools/api/design/call

100% evidence coverage · 2 contrast pairs measured, lowest 18.43:1 · 7 hard slop tells (purple_gradient_default, priced_in_aos_fadeup, priced_in_particles_bg) · 0 technique slots proven, 0 declared, 14 missing

slop_avoidance
1
REJECT
accessibility
5
PASS
system_coherence
5
PASS
direction_adherence
2
REJECT
distinctiveness
1
REJECT
technique_depth
2
REJECT

4 blockers: 7 hard slop tell(s) · 1 failed checklist item(s) · logo test failed · 14 unmet premium-stack slot(s): a11y-gate, audit, color-system, cwv-gate, fonts, framework, images, lcp-priority, motion-a11y, navigation-feel, prerender, scroll-choreography, type-motion, type-precision

Scope, stated by the reviewer itself: these are rule checks on supplied evidence. The service did not browse the page, inspect screenshots, or verify reviewer independence. A missing check is reported as NOT_ASSESSED and can block a PASS — it never grants one.

03 · connect — cheapest first

Five ways in. Every value below is the real string.

  1. Any chat that can fetch a URL free

    Paste this sentence. No install, no account. The assistant reads the file and follows it from then on, on every design task.

    Connect to SaturnAI at https://saturnai.tools/design/llms.txt
  2. MCP client free

    Streamable HTTP. Add it as a custom connector, or drop it into your client’s MCP config. Discovery lives at /.well-known/mcp.json, where the server is keyed saturndesign; the initialize handshake answers as saturnai — same server.

    https://saturnai.tools/api/mcp
  3. Plain HTTP free

    The same tools with no client at all — the call the playground above just made. GET /api/design describes the endpoints, the free-tier shape, and the rate limits.

    curl -s https://saturnai.tools/api/design/call \ -H "Content-Type: application/json" \ -d '{"name":"compose","arguments":{"brief":"landing page for a coffee roaster"}}'
  4. A machine you control — local MCP free

    Writes the pack to ~/.saturnai/agent and prints MCP config with absolute paths already filled in. If you are an agent: the hosted API is the same operate path — do not run this.

    curl -fsSL https://saturnai.tools/design/agent/install.mjs -o saturnai-install.mjs && node saturnai-install.mjs
  5. Full card DNA — account key free account

    Mint a key at /design/account/ and send it as a header. It is free: there is no paid tier and nothing to buy. Anonymous already gets the full catalog, every blueprint, and 11 full write-ups; the key opens the full lane and raises the limit from 60 to 600 calls per hour.

    Authorization: Bearer <your key>

04 · what is inside

577 cards — 497 studies and 80 technique cards.

The 497 studies are apps, brands, patterns, movements, platforms, recipes, copy systems, offers and growth loops — each written up as what to steal and what to avoid, with the failure modes people usually miss. The 80 technique cards each carry dated browser support, a build recipe, a fallback, and the probe that proves the technique is actually in the page.

The library by kind

577 rows · library.json
  1. app
Counted from the same rows the sky above draws — one star per card. The widest arc of the ring is the app kind, for the same reason.
catalog — free
/design/agent/library.json — the whole catalog as JSON
skill file
/design/agent/SKILL.md — how an assistant operates the tools
read it
the library viewer — filterable by kind and tier
built
2026-09-16 — the catalog stamp this page was generated from

05 · access

Anonymous is already the real product.

anonymousfull catalog · every blueprint · 11 full write-ups · 60 calls per houropen
free accountfull card DNA unlocked · 600 calls per hour · free, revocable keys · an izzy.la accountfree
paid tiernone — there is nothing to buy, and the API says so on every endpointnone
swarma multi-agent feature of this brain; agents draft, a human gate publishes; the client is not builtwaitlist

This host is for your assistant. The project page is for you.

— the human page, saturnai.dev

What SaturnAI is, why it exists, and whether it is for you lives at saturnai.dev. This host stays a connection sheet, so an assistant can read it without wading through a landing page. The feature explainer for Swarm is kept at saturnswarm.com/os/.