DocMCP – Index any docs site locally, search it from Claude via MCP

Name: DocMCP – Index any docs site locally, search it from Claude via MCP
Availability: InStock
Author: pieeee

by pieeee·Mar 5, 2026·2 points·0 comments

Visit Project View on HN

AI Analysis

●●●BangerBig BrainSolve My Problem

Hybrid BM25 + vector search via MCP beats pure keyword or pure semantic for API docs.

Strengths

•Hybrid search approach (FTS5 BM25 + embeddings + RRF) genuinely solves the "flex-wrap vs how to make elements wrap" semantic-vs-exact tradeoff that pure solutions miss.
•All data stays local; embedding providers optional (can run zero-API-cost BM25-only).
•Tight MCP integration works across Claude Desktop, Claude Code, and Cursor without friction.

Weaknesses

•Crawl-and-index workflow means docs must be updated manually or via cron; real-time doc updates aren't automatic.
•No comparison to Retrieval Augmented Generation (RAG) built into Cursor or baseline Claude training; unclear if hybrid search justifies setup vs copy-paste.

Post Description

Kept running into the same problem while coding with Claude: library docs are either outdated in its training data or I'm copy-pasting pages into the chat. Built DocMCP – an MCP server that crawls documentation sites, chunks and indexes them locally in SQLite, and exposes a search tool Claude can call directly. The interesting bit is the search: pure keyword search misses semantic queries ("how to make elements wrap" → flex-wrap), and pure vector search misses exact API names. So it does both – BM25 via FTS5 for exact terms, vector embeddings for semantics – and merges results with Reciprocal Rank Fusion. All data stays local in ~/.docmcp/. Embedding providers are optional (Anthropic Voyage, OpenAI, or BM25-only with no API key). Setup is `npm install -g @pieeee/docmcp` then `docmcp add https://react.dev`. Works with Claude Code, Claude Desktop, and Cursor. GitHub: https://github.com/pieeee/docmcp