Grounded Docs MCP Server 🔌 MCP server Open source

What it does

Grounded Docs, published as Docs MCP Server, builds a private index of the library documentation your project depends on so your coding agent can look up current, version-specific answers instead of guessing. You point it at documentation websites, GitHub repositories, npm or PyPI packages, local folders or zip archives. It crawls them, splits the content into chunks and stores them locally. Your agent can then search a particular library at a particular version. It handles HTML, Markdown, PDF, Office documents, notebooks and source code, and it uses llms.txt when a site publishes one. Search works on keywords alone, and you can add an embedding provider for semantic search. It is a self-hosted alternative to services such as Context7 and Ref.

Tools it gives your agent

• scrape_docs and refresh_version index a library, or re-index one
• search_docs, list_libraries and find_version query the index
• fetch_url fetches a single page as Markdown
• list_jobs, get_job_info, cancel_job and clear_completed_jobs manage indexing jobs
• remove_docs deletes a library from the index

Works with

It runs on your own machine through Node.js or Docker. You can launch it over stdio or run it as a long-lived server with SSE and Streamable HTTP endpoints plus a web UI on port 6280. The docs include setup for Claude Desktop, Cursor, Windsurf, Zed, and Cline and Roo in VS Code. It also ships a CLI and Agent Skills for agents that prefer shell commands.

How to connect

Start the server and add documentation through the web UI, then connect your client:

npx @arabold/docs-mcp-server@latest

{
  "mcpServers": {
    "docs-mcp-server": {
      "type": "sse",
      "url": "http://localhost:6280/sse"
    }
  }
}

Cost and limits

Free under the MIT licence and requires Node.js 22+. Semantic search is optional but recommended. It needs an embedding provider, such as OpenAI with your own API key or a local Ollama model, and each provider has its own cost.

Safety notes

The server fetches remote URLs and can read local files. By default it blocks private network targets and limits file access to allowed roots, so keep those defaults unless you trust the host. If you expose it beyond localhost, turn on its OAuth2 authentication. Anonymous usage telemetry is sent to PostHog, and you can turn it off with --no-telemetry or DOCS_MCP_TELEMETRY=false.