Docs MCP Server

Docs MCP Server attacks the most persistent failure mode in AI-assisted coding: the model confidently writing an API that was removed two major versions ago. It indexes documentation from wherever it actually lives — a docs website, a GitHub repository, an npm or PyPI package, or files on your disk — and serves it back over the Model Context Protocol scoped to the exact version your project depends on.

The differentiator is that you own the index. Context7, Nia and Ref.Tools solve the same problem as hosted services, which means your queries and often your code context leave your machine, and coverage is whatever the vendor has crawled. Here the pipeline runs locally: you point it at the sources you care about, including private or internal documentation that no hosted crawler will ever see, and nothing is transmitted except to whichever embedding model you choose.

Ingestion is broad — over 90 source-code languages plus PDF, Word, Excel and Markdown documents — so internal design docs and vendor PDFs can sit in the same searchable index as your dependency docs. Embeddings are pluggable across OpenAI, Google Gemini and local Ollama models; the Ollama path keeps the entire stack on your own hardware with no API key and no per-query cost.

Run it with npx @arabold/docs-mcp-server@latest for a quick trial, or the ghcr.io/arabold/docs-mcp-server container with mounted data and config volumes for a persistent index shared across a team. It exposes an HTTP interface on port 6280 alongside stdio, and works with Claude, Cline, Cursor and other MCP clients. MIT-licensed.