Watch a coding agent start a task and you’ll see the same opening every time: list the directory, grep for a symbol, open six files, work out how they relate. It is thorough, it is slow, and you pay for all of it — again on the next task, in the same repository.

Graft does that exploration once. It parses the repository with tree-sitter into a graph of markdown nodes — subsystems, APIs, concepts — with typed links between them, and wires the agent to read the graph instead of the files. The project reports 3× faster and 4× cheaper sessions, and on SWE-bench Verified 66% correctness against a 54% baseline with 23% fewer tokens. Treat those as the maintainers’ benchmark rather than a guarantee; the mechanism is sound either way.

Install 

  npm install -g @nanonets/graft
cd ~/projects/my-app
graft init      # asks which agents to wire, then builds the graph

init writes a graft/ directory and adds it to .gitignore — the graph is a cache, regenerable from source, not something to commit. For Claude Code it also installs a skill at .claude/skills/graft/SKILL.md without overwriting your existing settings.

The commands

CommandWhat it does
graft buildBuild the structural graph with tree-sitter — no API key, no cost
graft build --deepAdd LLM-written summaries and concept nodes
graft ask "<task>"Return the nodes and files relevant to a task
graft grep "<regex>"Search, grouped by enclosing symbol rather than by line
graft mapDirectory clusters, hubs, and hotspots — orientation in a new repo
graft checkReport where the graph has drifted from the code
graft vizOpen the graph in an interactive view

The split that matters: build is free and deterministic, --deep calls a model. You can run Graft indefinitely without an API key and still get the structure; the deep pass buys prose summaries on top.

graft ask is worth running yourself, not just letting the agent call it. It is the fastest way to answer “where does this actually live” in a repository you didn’t write.

With Claude Code

Choose Claude Code during init and you get context injection on each prompt, a statusline showing graph size and staleness, an automatic structural refresh after edits, and a blast-radius view of what depends on the file you just changed. Graft also registers an MCP server exposing the same operations as tools — graft_find_code, graft_file_api, graft_trace_calls, graft_repo_map, and friends — so any MCP-capable agent can use the graph.

Language coverage

FidelityLanguages
Full, scope-aware with cross-file resolutionTypeScript/JavaScript, Python, Go, Java
Symbols and generic call edgesRust, C, C++, C#, Ruby, PHP, Kotlin, Scala, Swift, Elixir, Dart, and more
Compiler-grade, opt-in --lspAnything with a language server — rust-analyzer, clangd, gopls, pyright

Monorepos build one graph and scope answers per package (graft ask "..." --in apps/api/); submodules come along with --follow-submodules.

Worth knowing before you adopt it

  • A stale graph is worse than none. Run graft check after a big merge; the auto-sync covers ordinary edits, not a rebase that moved half the tree.
  • Deep mode costs tokens. Building summaries for a large repository is a real bill. Start structural, add --deep only if the agent still gets lost.
  • Telemetry is on by default. Anonymous usage buckets, no code or paths, and graft telemetry disable or DO_NOT_TRACK=1 turns it off.
  • MIT licensed, and the graph lives in your repository — nothing to migrate if you stop.

Next

For the same idea as an MCP server with a database behind it → Codebase Memory MCP

Edit this page

Last updated 23 Aug 2026, 00:00 UTC. history