Codebase Memory MCP

Codebase Memory MCP gives AI coding agents a persistent map of a repository's structure, so they can answer questions about dependencies and call chains without repeatedly reading whole files. It's for developers using Claude Code, Codex CLI, Cursor or other MCP-compatible clients who want their agents to spend less context on code exploration.

The server runs locally on macOS, Linux and Windows. It's open source under the MIT license and needs no hosted service or API key. It has no embedded LLM; your coding agent supplies the reasoning. The server's parsing and search stay on your machine, while your chosen agent handles its own model connection. The native executable doesn't need Docker or a language runtime.

Its graph connects functions and classes with HTTP routes and links between services. Tree-sitter parsing covers languages including Python, TypeScript, Go, Rust and Java. Hybrid LSP type resolution adds information about imports, inheritance and return types to help resolve calls across files. Agents can trace data flow, identify unused functions, inspect architecture and assess which symbols an uncommitted change affects.

Semantic search uses bundled nomic-embed-code embeddings on-device to find code by meaning. Clone detection finds near-duplicate functions, and the graph can connect multiple repositories in one store. A background watcher updates the index as files change. Teams can share a compressed graph snapshot, keep architectural decisions alongside it, and explore nodes and relationships through an optional local 3D browser view.