Reference
Docs for AI tools (MCP)
@matrajs/mcp is this documentation as a
Model Context Protocol server. Connect it and an
AI assistant answers questions about Matra from these pages rather than from memory — which matters
for a framework whose API is inferred from an array and has no ProseMirror underneath, because
memory will confidently describe something else.
npx -y @matrajs/mcp # stdio · what a desktop client spawns
npx -y @matrajs/mcp --http # http://localhost:3333/mcpStep 1 · Connect it
Claude Code
claude mcp add matra -- npx -y @matrajs/mcpClaude Desktop · in claude_desktop_config.json
{
"mcpServers": {
"matra": { "command": "npx", "args": ["-y", "@matrajs/mcp"] }
}
} Cursor · the same object under "mcpServers" in .cursor/mcp.json.
Codex · in ~/.codex/config.toml
[mcp_servers.matra]
command = "npx"
args = ["-y", "@matrajs/mcp"] Anything that speaks HTTP · run npx -y @matrajs/mcp --http 3333 and point the client at http://localhost:3333/mcp.
Step 2 · Ask
Ask the assistant how to add search and replace, or what ctx.mark() is for. It calls
search_docs, reads the page it needs with read_doc, and quotes it.
Nothing is fetched at runtime · every page ships inside the package.
What it serves
| Tool | Does |
|---|---|
list_docs | Every page, with its slug and a one-line description. |
read_doc { slug } | One page, as Markdown. |
search_docs { query, limit? } | Ranked pages with a snippet each. |
Every page is also a resource at matra://docs/<slug>. The pages are the
repository's Markdown — the README, the engine notes, benchmarks, security, the changelog —
and every page of this site, converted to Markdown when the package is built.
From code
import { createServer } from '@matrajs/mcp'
const server = createServer(docs)
server.handle({ jsonrpc: '2.0', id: 1, method: 'tools/list' }) createServer is the protocol without a transport: one message in, one reply out.
Put it behind whatever transport you already run.