The problem it names is real: good servers exist but finding them means trawling GitHub, awesome-lists and forums, and an agent hits that wall mid-task. Precomputed embeddings are the design decision that makes it usable — the lookup is sub-second because the work happened at build time, not at query time. Note that it recommends servers; installing and configuring the one it finds is still yours to do.
A meta-server whose job is discovery. When an agent needs a capability it does not have, it describes the need in natural language and this returns candidate MCP servers from an aggregated database.
- 2,874+ MCP servers in the bundled database, aggregated from the official modelcontextprotocol/servers list and the Punkpeye and Appcypher community collections, with deduplication across sources
- Semantic search over precomputed embeddings, so a plain description like "check domain availability" resolves without keyword matching
- The database refreshes automatically every 3 hours from the upstream sources
- Stdio for editor clients, or HTTP with configurable host and port for development
Python 3.13 or newer. Run it as `uvx mcp-mcp` with no configuration, or add it to Claude Code in one command. Nothing to authenticate against; the index ships with the package and the embeddings are precomputed, so first startup is fast rather than an indexing job.
One command — uvx mcp-mcp