Missed the latest GitHub check; these facts may be older.
Overview
Ref MCP exposes two tools, ref_search_documentation and ref_read_url, so a coding agent can search public docs, GitHub, and private repos or PDFs, then read pages as markdown. It uses MCP session history to drop repeated search results and trims fetched pages to the most relevant ~5k tokens. This repo holds the legacy stdio server; the recommended streamable-HTTP endpoint is hosted at api.ref.tools and needs an API key.
Who it is for: Developers giving Claude Code or Cursor agents up-to-date docs lookup
Strengths
Session-aware search never repeats earlier results, so agents can page and refine in one call
Page reads return the most relevant ~5k tokens instead of 20k+ raw fetches
Runs over stdio via npx or hosted streamable HTTP; Cursor install links provided
Offers OpenAI deep research tool names (search, fetch) alongside its own
Weaknesses
Requires a Ref API key from signing up; search depends on the hosted service
Repo contains only the legacy stdio server, not the recommended HTTP one
Only two tools, limited to documentation search and URL reading