Your agent reads garn natively.

Wire up the MCP server

garn-mcp is a standalone package — no clone, no build. It bundles a metadata snapshot, so it runs straight from npx. Register it in your AI client:

claude mcp add garn -- npx -y garn-mcp

Cursor and other clients take the same command as a JSON entry (in .cursor/mcp.json, or your client's MCP config):

{ "mcpServers": { "garn": { "command": "npx", "args": ["-y", "garn-mcp"] } } }

Six tools cover the judgment layer:

• garn_searchFind components by keyword + pattern/category filters. Ranked. Start here.
• garn_viewFull reference for one component; `sections` slices it to save tokens.
• garn_get_examplesType-checked example code, filterable by role (e.g. canonical).
• garn_match_recipeMap an intent → best-fit compositions, each with its a11y contract.
• garn_relationsWhat pairs with, composes with, or replaces what.
• garn_auditLint a composition against a11y contracts, token discipline, and do/don'ts.

Developing garn itself? Inside this repo, Claude Code picks the server up from the root .mcp.json automatically — it serves the live in-repo data layer via pnpm -s --filter garn-mcp start, so run pnpm info:build once first to generate it.

No MCP? Read the corpus

The same metadata is served flat for tools that just read text: /llms.txt (the index) and /llms-full.txt (the full corpus — every component's usage, props, contracts, and recipes). Copy-in consumers also get a generated AI rules file in their repo via garn init.