Working with AI
Kimen treats agent legibility as a deliverable: an API a capable agent cannot wire correctly from its description alone is treated as a defect. Everything below is generated from the component sources and shipped with the package, so it can never drift from the code.
llms.txt
Section titled “llms.txt”The library, indexed for language models the way the llms.txt convention describes. The index is short — what Kimen is, how to install and load it, and one line per component saying when to use it — and each line links one Markdown document with that component’s complete contract: when-to-use / when-not-to-use guidance, typed attributes (required, bounded and URL attributes marked), slots, CSS parts, events, methods, styling tokens and curated runnable examples.
- On this site:
/llms.txtis the index. Every document it links is served under/llms/— for example/llms/ki-button.md— and/llms-full.txtis all of it in one file. - In the npm package:
@kimen/elementsshipsllms.txtwith relative links tollms/*.mdandllms-full.txt, so an agent reads the documents of the version that is installed, offline. - In the repository:
llms.txtat the repository root.
Point your tool at the index — @docs in Cursor or Windsurf, or paste the
URL into a ChatGPT/Claude conversation — and the agent reads about two
thousand tokens, then fetches only the components it decides to use.
Skills and the MCP server
Section titled “Skills and the MCP server”To set up a coding agent rather than point it at files, install the Kimen
agent skills (kimen-components and kimen-genui) and connect the
kimen-mcp server, which lists and describes the catalog, validates UI
specs and renders them as live components in MCP Apps hosts. Setup for
Claude Code, Claude Desktop, Codex, Antigravity, OpenCode and other clients:
Connect your agent.
custom-elements.json
Section titled “custom-elements.json”The Custom Elements Manifest is the interchange format the ecosystem reads: Storybook, IDE completions (VS Code custom data, JetBrains web-types) and this site’s own API tables all consume it.
- In the npm package: the
customElementsfield of@kimen/elementspoints togenerated/custom-elements.json. - On this site: served at
/assets/elements/custom-elements.jsonfor tools that want to fetch it.
Kimen extends the standard schema with whenToUse / whenNotToUse fields
per component — the same guidance rendered on each component page.
The catalog: from description to guardrail
Section titled “The catalog: from description to guardrail”Reading about components is half the story; emitting UI safely is the
other half. @kimen/catalog gives agents a validated output format — the
UI spec — and gives hosts a renderer that
enforces it fail-closed. For protocol-level integration, Kimen ships
adapters for A2UI and
MCP Apps, where a Kimen surface reaches an
MCP host as a self-contained ui:// resource.
To configure the model itself, @kimen/emitter derives the guidance
artifacts from the catalog in use: a system prompt carrying each
component’s usage guidance verbatim, a catalog-specialized JSON Schema
with provider-strict lowerings, and a provider-neutral tool definition —
plus the normalize-and-repair helpers that close the reliability loop. See
Emitting specs with any LLM.
llms.txt tells a model what exists; the catalog constrains what it may build; the emitter teaches any model to build it. Shipping all three is the point.