Ref Tools Mcp

This server tackles a real problem with AI coding agents: they often pull in way too much documentation context, which both costs tokens and degrades model performance through context rot. It provides two tools for searching and reading technical docs, both public and private. The smart bit is how it handles sessions: repeated searches filter out already-seen results, and when reading a page it uses your search history to return only the relevant 5k tokens instead of dumping 20k+ tokens of an entire docs site into context. Works over stdio or HTTP transport, and the tools automatically adapt their names when used with OpenAI's deep research client.