DevDocs-MCP is a local MCP server that provides documentation authority for AI agents, based on the DevDocs.io documentation dataset, and is designed to eliminate AI hallucinations and provide version-aware and deterministic documentation.
[Purpose] DevDocs-MCP acts as middleware between AI agents and documentation sources, solving three major problems faced by modern AI programming agents: hallucinations (guessing API signatures or using deprecated methods), latency (fetching online documentation for every query is slow), and version drift (confusing React 18 and React 19 features). It provides "just-in-time" documentation for AI assistants by retrieving documentation on demand from local storage and filtering it according to project-specific dependency versions.
[Main capabilities]
- Offline-first: documentation is cached locally, and no internet is needed after ingestion.
- Ranked fuzzy search: quickly find relevant entries (such as "intersection observer").
- Version-aware: automatically maps to specific library versions in the project.
- Pure Node architecture: uses sql.js to achieve zero native dependencies (no Python/C++ build step required).
- Structured output: returns clean, LLM-optimized content.
- Project-aware context manager: pins documentation to package.json versions.
- Lazy ingestion engine: caches DevDocs offline.
[Usage]
- Environment requirements: Node.js 18+ (tested on 18.x and 20.x).
- Clone and install: use pnpm for development and contributions (strictly recommended).
- Configure the environment: in the .env file, DEVDOCS_VOLUME_SOURCE supports 3 storage modes: named volume (default, Docker-managed), host path (such as ./my-docs-storage).
- Start the server: docker compose up -d or run the public image directly.
- The unified mount point /app/data contains mcp.db (SQLite registry) and docs/ (cached documentation files).
- Configure storage in .env and then start; each time it can point to a new volume or an existing volume.
- Ingest documentation: after the server connects to the agent, use the ingest tool to download documentation for a specific tech stack; the agent will handle it automatically, list the available tools, and make decisions.
- AI agent configuration: RooCode/Cline/Claude Desktop recommendsthe STDIO method, adding configuration in mcp_settings.json, setting cwd to load .env and using relative paths; remote/web clients can use HTTP SSE, starting with --port 3000, and the client configures the SSE URL http://your-server:3000/mcp/sse，类型 SSE.
[Applicable scenarios] Suitable for development scenarios using AI programming agents, especially local development environments that require precise API documentation, avoid version confusion, are network-constrained, or pursue low latency. Supports multiple agents and clients.
[Limitations]
- Project status: actively and heavily under development; existing features still need testing and more new features still need to be implemented (such as optimal caching, automatic updates to the latest documentation and on-demand updates, optimal querying, etc.). Design decisions, APIs, and structure are expected to evolve based on community feedback.
- Not an AI agent or code generator.
- Not a web crawler (uses the structured DevDocs dataset).
- Not a replacement for the DevDocs interface.
- Not a hosted SaaS (100% local operation).
- Development and contributions strictly use pnpm.
- The DevDocs fetch URL is defined in src/config/constants.ts; if the community changes the path, the configuration needs to be updated.
- No specific pricing, cost, platform, or performance data is officially provided.
- License: MIT.
