Quilt

Design agent

design.mp4

Publish Agent

browser-use.mp4

Marketing artifacts, brand-locked, published from your own browser.

A local desktop studio. Give it a brief; it produces a landing page, an email, a poster, a set of social cards, checks them against a design linter and a weighted jury, and then deploys them into the tools you already use and goes back to confirm they really arrived.

   BRAND      AUDIENCE     BRIEF       BUILD        VERIFY      PUBLISH
   -----      --------     -----       -----        ------      -------
   scrape --> CSV      --> kernel  --> agent    --> lint    --> replay a
   vision     data agent   one         8 tools      repair      recorded path
   extract    proposes     headline    1 render     jury        or drive Chrome
   tokens     the rubric   one CTA                              |
   + voice    you approve  shared                               v
              it; code     verbatim                         independent
              scores it                                     verification

The audience lane is optional. Skip it and one brief produces one set for the market; use it and the same brief produces one set per company, merged against the columns you imported.

Why it exists

Three claims, all load-bearing:

1. Generation costs tokens. Refinement must not. Every artifact satisfies a structural contract, so nudging a heading, swapping a token colour or reordering two cards is a typed mutation applied in 18ms with no model call. See docs/DECISIONS.md 005.

2. Success has to be proven, not claimed. An agent reporting done and an API returning 200 are both claims. Every publication is re-checked in a context the publisher never touched, and the evidence has to come from the artifact rather than from a caption the agent typed itself. See docs/DECISIONS.md 009.

   Taking one back is held to the same standard, and it is harder: absence cannot be proved by failing to find something, so a retraction has three outcomes rather than two and "we could not look" is never success. See docs/DECISIONS.md 036.

3. Who it is for is a rubric, not an opinion. Accounts come in as a CSV from whatever CRM you already have and are tiered by weighted conditions over their own columns, deterministically, with the reasons attached to every tier. A merge field that does not resolve stays visible in the artifact rather than becoming an empty string. See docs/DECISIONS.md 037 and 038.

   A data agent drafts that rubric by reading your actual columns and their actual values, and then stops: it proposes, you edit, and arithmetic scores. Judgement is the model's job and consistency is the code's. See 039.

Running it

Requires Node 24 and pnpm 11.

pnpm install
pnpm dev          # server on :7457, web on :7456
pnpm app          # the packaged Electron app
pnpm dist:mac     # build a signed-ish .app

Add a provider key in Settings. Anthropic, OpenAI and MiniMax are supported; a key is stored AES-256-GCM at rest, mode 0600, and never leaves the server.

pnpm test         # 813 tests, no key or network required
pnpm typecheck
pnpm eval         # real briefs through the real pipeline. Costs money.

Layout

  apps/server     Hono on 127.0.0.1, SQLite, Playwright, the agents
  apps/web        React 19, Zustand, Tailwind 4, a 142-line hand-rolled router
  apps/desktop    Electron shell: menu, window state, native context menus
  packages/shared events, model capabilities, the one product-name constant

  craft/          13 design rulebooks. Rules tagged [lint] are enforced by code
  recipes/        14 artifact types. Frontmatter + prose, user-overridable
  skills/         publish destinations. One path through one product surface
  evals/          real briefs, scored, diffed against a committed baseline
  docs/           why every non-obvious thing is the way it is

craft/, recipes/ and skills/ shadow from ~/.quilt/, read at request time. A platform UI the agent cannot navigate is fixed by editing a markdown file, not by shipping a release.

Docs

File	What it answers
docs/PLAN.md	what the product is, surface by surface
docs/DECISIONS.md	why each choice beat its alternative, numbered and cited from the code
docs/CRAFT.md	how quality is measured and repaired
docs/PROVIDERS.md	the verified facts and the twelve traps behind every capability gate
evals/README.md	how to write and run an eval case

What it deliberately is not

Not a cloud service, not a chat product, not a general coding agent, not a template gallery. Each of those is a decision with a record.