Graft 🔌 MCP server Open source

What it does

Graft builds a local, linked-markdown knowledge graph of your codebase (via tree-sitter parsing plus optional LLM summarization) so coding agents can look up structure and context instead of re-exploring the repo from scratch on every task. Unlike a one-off grep or a static docs site, the graph tracks working-tree state, including uncommitted edits, and regenerates automatically as code changes, so it stays current mid-session.

Tools it gives your agent

• A CLI (graft) to build, initialize and maintain the graph
• An MCP server for native tool access from MCP-capable clients
• Deep, first-class integration with Claude Code, including a statusline and auto-sync
• Support for 23 languages, including TypeScript, Python, Go, Java, Rust and C++
• Optional LLM-generated summaries layered on top of the tree-sitter structural data

Works with

Claude Code (deepest integration), plus Cursor, Codex/OpenCode, Gemini, GitHub Copilot, Windsurf, Grok (xAI), Kiro and AdaL through its MCP server.

How to install

npm install -g @nanonets/graft
graft init

Can also be run without a global install via npx @nanonets/graft init.

Maintenance and safety

Actively maintained, MIT licensed, with 9k+ stars and 800+ forks. The graph is stored locally and is regenerable (not meant to be committed to your repo); telemetry is optional and anonymous and can be disabled. No code or file paths are sent anywhere without an explicit LLM-summarization request using your own provider key (OpenAI, Anthropic, OpenRouter, Fireworks, etc.). Treat the README's benchmark numbers ("up to 4x cheaper, 3x faster," SWE-bench gains) as vendor-reported rather than independently verified.

Who should use it

Developers working in large or unfamiliar codebases with Claude Code, Cursor, or similar agents who want to cut down on repeated file exploration and token spend by giving the agent a persistent, auto-updating map of the codebase's structure.