GraftMCP Server

Every task, a coding agent starts blind. It greps a term, opens a file, follows an import, backs out and tries again — paying for the same reconnaissance on every single request. Graft is an open-source context layer that does that exploration once, ahead of time, and keeps the result.

It builds a per-symbol code graph of the repository as a folder of linked markdown files: summaries, "crux" excerpts extracted with tree-sitter, and the call and import edges between them. Because the cache is plain markdown on disk, it is regenerable, diffable and readable by a human, and it stays in sync with the working tree rather than the last commit — uncommitted changes are included. Twenty-three languages are supported, among them TypeScript, Python, Go, Java and Rust, and the layout handles monorepos, submodules and multi-repo folders.

The MCP server exposes six tools — graft_find_code, graft_file_api, graft_trace_calls, graft_find_all, graft_repo_map and graft_check_freshness — so an agent can ask where something is defined, trace who calls it, or pull a map of the repository without reading the files. graft init also wires the graph directly into Claude Code and other agents, and an interactive dependency visualisation ships alongside.

The project publishes its own benchmark numbers: on SWE-bench Verified, correctness moving from 54% to 66%, with 42% token savings and 46% fewer tool calls, and up to 4× cheaper and 3× faster execution against its baseline. LLM provider support is vendor-neutral. Distributed on npm as @nanonets/graft under MIT, and actively released — the package has shipped continuously since July 2026.