Runs a local semantic code search engine that exposes three distinct MCP tools: search_code for finding definitions and structure via BM25 and dense vectors, search_docs for architecture queries, and search_history for Git-based debugging. Uses sentence-transformers for embeddings so everything stays offline with no API keys. Supports full AST parsing for Python, JavaScript, TypeScript, Java, Go, Rust, C/C++, Ruby, and Kotlin, with fallback indexing for other languages. Respects .gitignore automatically. The main win is precise retrieval instead of dumping entire files into context, which the author claims cuts token usage by half. Can run as stdio per project or as a shared SSE server to avoid loading the 600MB model multiple times across workspaces.
claude mcp add --transport stdio kapillamba4-code-memory uvx code-memory