mcp-server-dev-docs

An MCP server that provides live, clean, LLM-ready markdown documentation for popular fullstack technologies — and any custom documentation URL.

It fetches pages through https://r.jina.ai/ under the hood to strip navigation, ads, and boilerplate, returning clean markdown. No API key or authentication required.

Features

• Built-in presets for React 19, Next.js, TypeScript, Tailwind CSS, Zustand, TanStack Query v5, Node.js, Prisma, and Express.

• Fetch any arbitrary documentation URL as clean markdown.

• Find documentation website URLs by technology (and optional version) via DuckDuckGo HTML, then return clean markdown for the top results through Jina Reader.

• Detect dependency versions from the current project's package.json.

• Automatic truncation of very large pages (> 25,000 characters) to keep responses LLM-friendly.

• Zero configuration and no API keys.

Related MCP server: Context7 MCP

Tools

list_supported_presets

Lists the built-in preset keys, their target URLs, and descriptions. Takes no arguments.

get_fullstack_docs

Fetches clean markdown documentation. Provide exactly one of:

detect_project_versions

Reads package.json from the current working directory (process.cwd()) and returns a JSON summary of dependencies and devDependencies with their declared versions. Takes no arguments.

find_docs_urls

Finds documentation website URLs for a technology via DuckDuckGo HTML, then returns clean markdown for up to the top 2 results (fetched in parallel via Jina Reader).

URL discovery behavior:

1. Primary query (when the technology has a known docs domain):
   site:<domain> <technology> <version?> <query>
   Example: site:react.dev react 19 useActionState

2. Fallback if the primary query returns no links:
   <technology> <version?> docs <query>

3. Extract up to two result URLs, fetch markdown for both with Promise.all, and return them as:

--- Result 1: <url> ---

...

--- Result 2: <url> ---

...

Mapped domains include react → react.dev, nextjs/next → nextjs.org, tailwind/tailwindcss → tailwindcss.com, and zustand → zustand-demo.pmnd.rs.

Note: DuckDuckGo HTML may rate-limit or challenge automated searches after many requests. If URL discovery fails intermittently, wait and retry, or use get_fullstack_docs with a preset/customUrl.

Supported presets

Installation & Usage

The server runs over stdio and is intended to be launched by an MCP client. The easiest way is via npx (no global install needed):

npx mcp-server-dev-docs

Cursor

Add the server to your Cursor MCP config at ~/.cursor/mcp.json (global) or .cursor/mcp.json (per-project):

{
  "mcpServers": {
    "dev-docs": {
      "command": "npx",
      "args": ["-y", "mcp-server-dev-docs"]
    }
  }
}

Then reload Cursor. The tools list_supported_presets, get_fullstack_docs, detect_project_versions, and find_docs_urls will be available to the agent.

Claude Desktop

Add the server to claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "dev-docs": {
      "command": "npx",
      "args": ["-y", "mcp-server-dev-docs"]
    }
  }
}

Restart Claude Desktop to load the server.

Local Development

Requires Node.js >= 18 and Yarn.

git clone https://github.com/Orel-Toledano/mcp-server-dev-docs.git
cd mcp-server-dev-docs
yarn install
yarn build      # compiles TypeScript to ./build and marks the entry executable
yarn start      # runs the compiled server over stdio

To point an MCP client at your local build, use an absolute path:

{
  "mcpServers": {
    "dev-docs": {
      "command": "node",
      "args": ["/absolute/path/to/mcp-server-dev-docs/build/index.js"]
    }
  }
}

Smoke test

You can exercise the JSON-RPC handshake without a full client:

printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"1.0.0"}}}' \
  '{"jsonrpc":"2.0","method":"notifications/initialized"}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}' \
  | node build/index.js

Or use the MCP Inspector:

npx @modelcontextprotocol/inspector node build/index.js

Publishing

This package builds automatically before publish via the prepare script.

# Inspect the tarball contents first (dry run)
yarn pack

# Authenticate with the npm registry
yarn login

# Publish (public access for unscoped/public packages)
yarn publish --access public

Only the compiled build/ directory is published (see the files field in package.json).

How it works

• get_fullstack_docs resolves a target URL (from a preset or customUrl) and requests https://r.jina.ai/<target-url> with X-Return-Format: markdown.

• find_docs_urls discovers candidate documentation URLs via DuckDuckGo HTML, then fetches each through the same Jina Reader path.

• Large pages are truncated at 25,000 characters to keep responses LLM-friendly.

Project structure

src/
  index.ts                 # MCP server entry (stdio)
  presets.ts               # Built-in documentation presets
  fetchDocs.ts             # Jina Reader fetch + truncation
  search.ts                # Domain map, DDG query/parse helpers
  tools/
    index.ts               # Registers all tools
    listSupportedPresets.ts
    getFullstackDocs.ts
    detectProjectVersions.ts
    findDocsUrls.ts

License

MIT