ref-tools-mcp

Why we included this project

Coding agents like Claude Code or Cursor spend a surprising share of their context window fetching library documentation, and much of what they pull in is noise. Ref works around that by serving docs through an MCP server with two tools: a search that returns ranked results, and a reader that trims a fetched page down to what the agent actually needs, based on what it has already looked for in the session. Repeated lookups get deduplicated, so the agent refines its query instead of re-fetching the same pages. If you have watched token bills climb while an AI pair programmer digs through public APIs, private repos, or PDF docs, this is aimed directly at you. It installs cleanly into Cursor or Claude Code over stdio or HTTP, so it feels like a practical add-on rather than a research experiment.