Multi-repo code-search engine

Coding agents are only as good as the context they can find. Point one at a single file and it does fine; point it at a monorepo with a dozen services and it starts guessing, because grep doesn’t understand code and most semantic search tools only index one repository at a time.

This is a multi-repo semantic code-search MCP server, written in Rust, that indexes many repositories at once and exposes hybrid retrieval — vector search plus BM25, tree-sitter AST chunking so results respect function and class boundaries instead of arbitrary line windows, and DuckDB underneath for storage and querying. It runs fully offline: no code or embeddings leave the machine.

It speaks the Model Context Protocol, so it plugs into OpenCode, Claude Code, Cursor, or any other MCP client without custom glue per editor. In day-to-day use it’s the difference between an agent that has to be told exactly where something lives and one that can just be asked.

A deeper video series walking through the retrieval design (why hybrid, why AST chunking, why DuckDB) is planned — this page will link to it once it’s out.