Using llms.txt with AI tools
There are two ways to give an AI tool helix-ui: the MCP server (best — live tools) and the static llms.txt endpoints (works with any tool that can fetch a URL).
Fastest path: the MCP server
If your tool speaks the Model Context Protocol (Claude Code, Cursor, Claude Desktop, Windsurf…), connect the helix-ui MCP server. It exposes live tools — helix-ui.search, helix-ui.validate, helix-ui.fix_jsx, helix-ui.express_dna — and the manifests as resources, all backed by the same artifacts below.
# Claude Codeclaude mcp add helix-ui -- npx -y @helix-ui/mcp
# Anything else: run it directly over stdionpx -y @helix-ui/mcp// or add to your mcp.json / client config{ "mcpServers": { "helix-ui": { "command": "npx", "args": ["-y", "@helix-ui/mcp"] } }}Static endpoints
helix-ui publishes static endpoints with every build, discoverable via robots.txt, sitemap.xml, a <link rel="llms-txt"> tag, and an HTTP Link: header:
/llms.txt— a compact index of every doc page and every component with one-line descriptions. Carries a provenance line (schema, build id, date)./llms-full.txt— the full body of every doc page concatenated, plus the tokens manifest as JSON./components/<name>.md— one prompt-ready file per component./tokens-manifest.json,/components-manifest.json— machine-readable, each stamped with a$meta/schemaprovenance block so a client can cache-validate (ETag +If-None-Match304s are supported).
With Claude Code
Add helix-ui as a documentation source the agent can fetch:
# in your projectecho "Read https://your-helix-ui-host/llms.txt before suggesting any helix-ui component or token." >> CLAUDE.mdFor deeper context, swap in llms-full.txt. It is large but covers the whole system in one fetch.
With Cursor
Settings → Features → Docs → Add new doc → paste the URL of llms-full.txt. Cursor will index it and use it for completions.
What’s in llms-full.txt
- A short header that tells the model what helix-ui is.
- The full content of every page in
apps/docs/src/content/docs/. - Every component’s
spec.md(frontmatter + body), so the model sees props, accessibility notes, and example code together. - The tokens
manifest.json, embedded as a fenced JSON block.