A local RAG system that indexes your markdown documentation and exposes semantic search to Claude via MCP. Uses vector embeddings to retrieve relevant chunks instead of reading entire files, cutting token usage by roughly 40x. Provides tools for searching with directory and file pattern filters, indexing new documents, listing what's indexed, and reindexing when files change. Runs as a single binary with automatic GPU detection and works offline once models download. Useful when you're working with large doc repositories and want Claude to find relevant sections without burning context on full file reads. The CLI doubles as a standalone search tool if you need it outside of Claude sessions.