docs-mcp

An MCP server that lets AI agents search and read your documentation at query time, instead of relying on stale training data. Point it at a folder of Markdown and it exposes search_docs, get_doc and list_docs over stdio (for Claude Code, Cursor, Claude Desktop) and over Streamable HTTP (for a hosted, public endpoint like OpenAI's Docs MCP or the Microsoft Learn MCP Server).

Published as @apideck/docs-mcp. It runs as a standalone CLI, a Vercel function, or as a library inside an existing Node or Next.js site. The docs/ folder in this repo is both the sample content and the server's own documentation. Start there: docs/index.md.

pnpm install
pnpm audit -- --docs ./docs          # check the docs are complete, current, structured
pnpm search -- "vercel" --docs ./docs # try the index from the terminal
pnpm start -- --docs ./docs           # MCP over stdio
pnpm serve -- --docs ./docs --port 3000   # MCP over HTTP at http://localhost:3000/mcp

Build once (pnpm build) and the docs-mcp binary in dist/bin/ runs the same commands without tsx.

claude mcp add my-docs -- docs-mcp start --docs /path/to/docs --base-url https://docs.example.com

Or commit a .mcp.json next to the docs:

{
  "mcpServers": {
    "my-docs": {
      "command": "docs-mcp",
      "args": ["start", "--docs", "./docs", "--base-url", "https://docs.example.com", "--about", "the Example API reference"]
    }
  }
}

Shared flags: --docs, --base-url, --about, --name. Each falls back to DOCS_DIR, DOCS_BASE_URL, DOCS_ABOUT, DOCS_NAME.

• search_docs(query, limit?, path_prefix?): heading-level full-text search with title and heading boosts, prefix and fuzzy matching, and at most three hits per page.
• get_doc(path, section?): full page markdown with a header and section outline, or one section and its sub-sections. path also accepts a full URL.
• list_docs(path_prefix?, limit?): pages with title, description, word count and last-modified date.

Every page is also an MCP resource at docs://<path>. Full reference: docs/tools.md.

api/mcp.ts is a stateless Streamable HTTP function; vercel.json rewrites /mcp to it and bundles docs/** with the function. Set DOCS_BASE_URL and DOCS_ABOUT in the project environment and deploy. Details in docs/hosting.md.

Mount the handler inside a site that already builds its docs, so the endpoint lives next to them. A Next.js pages-router API route:

// src/pages/api/mcp.ts
import { createHttpHandler, DocStore } from '@apideck/docs-mcp'
import type { NextApiRequest, NextApiResponse } from 'next'
import path from 'path'

const store = new DocStore({ root: path.join(process.cwd(), 'public', 'md'), baseUrl: 'https://docs.example.com' })
const handler = createHttpHandler({ store, name: 'example-docs', about: 'the Example API documentation' })

export const config = { maxDuration: 60, api: { responseLimit: false } }
export default (req: NextApiRequest, res: NextApiResponse) => handler(req, res)

Add experimental.outputFileTracingIncludes: { '/api/mcp': ['./public/md/**/*'] } to next.config so the markdown ships with the function on Vercel. DocStore also takes a metadata(path) hook to supply titles, descriptions and canonical URLs from a build manifest, and get_doc accepts a full URL as well as a path. Pass extraTools to createServer/createHttpHandler to add host-specific tools (an API operation index, a coverage matrix, ...) alongside the three built-ins — see docs/tools.md. Exports: DocStore, createServer, createHttpHandler, createDocTools, toolResult, auditDocs, formatAuditReport.

pnpm typecheck
pnpm lint
pnpm test        # node:test via tsx, covers parsing, indexing, audit, MCP over in-memory and HTTP transports
pnpm build

Layout mirrors @apideck/mcp: src/ for the library, bin/ for the stricli CLI, api/ for the Vercel function, tests co-located as *.test.ts.

1. Every .md/.mdx/.markdown file under the docs root is read, frontmatter parsed, and the body split on ATX headings (code fences respected).
2. Each section becomes a document in a MiniSearch index with title, heading, content and path fields.
3. Queries run with all terms required first, falling back to any term, so a typo does not return nothing.
4. In start and serve, a recursive file watcher rebuilds the index 300 ms after the last change. On Vercel the index is built once per function instance and refreshed by each deploy.

MIT