Graft

Turbocharge Claude Code, Cursor, Codex, Gemini & every coding agent: faster, cheaper, with contextual understanding specific to your codebase.

Up to 4× cheaper and 3× faster, with better or no loss of correctness.

| Metric | Cold Claude Code | Claude Code with graft | |---|---|---| | Tool-call reduction | Baseline | +46% | | Token savings | Baseline | +42% | | Time savings | Baseline | +60% | | Correctness | 54% | 66% (+12 pts) |

Efficiency is a 162-run controlled benchmark (same agent, same file tools, only the context differs). Correctness is SWE-bench Verified, graded by the official harness — graft resolved 66% of instances tested against Cold Claude Code's 54%. [Efficiency method ↓](#benchmark) · [SWE-bench ↓](#swe-bench-verified) · [Per-repo numbers ↓](#tested-on-your-popular-repos)

---

Contents

• [Quick start](#quick-start)
• [The problem](#the-problem)
• [What Graft does](#what-graft-does)
• [Benchmark](#benchmark)
• [SWE-bench Verified](#swe-bench-verified)
• [How the graph gets built](#how-the-graph-gets-built)
• [Supported languages](#supported-languages)
• [What's in a node](#whats-in-a-node)
• [What runs where](#what-runs-where)
• [Agent integration](#agent-integration) — [MCP server](#mcp-server) · [Claude Code (deep integration)](#claude-code-deep-integration)
• [CLI](#cli)
• [Search & orient](#search--orient-graft-grep--graft-map) (graft grep / graft map)
• [Monorepos & multi-repo folders](#monorepos--multi-repo-folders)
• [Visualize it](#visualize-it-graft-viz) (graft viz)
• [Tested on your popular repos](#tested-on-your-popular-repos)
• [Development](#development)
• [License](#license)

---

Quick start

npm install -g @nanonets/graft   # install the CLI, once
graft init                       # build the graph + wire it into Claude Code

That is the whole setup. graft init asks which of your coding agents to wire up, builds graft/ from your code, and drops a statusline and hooks into .claude/, so from the next session on Graft rides along in Claude Code: it pulls the matching nodes into each prompt and rebuilds the graph in the background after every turn. No daemon, no re-indexing to remember, nothing to run or maintain by default — the graph is just files.

Nothing is written until you pick. Run graft init --dry-run to see every file it would touch first, or graft init --agents claude to skip the prompt and wire Claude Code alone.

graft build adds graft/ to your .gitignore automatically — the graph is a local, regenerable cache (like node_modules), not something you commit. What you share is the wiring init dropped into .claude/; each teammate runs graft build to generate their own graph:

git add .claude && git commit -m "wire in graft"

Prefer not to install globally? npx @nanonets/graft init works the same way.

---

The problem

Every task, your coding agent starts blind. Before it changes anything, it re-explores the repo: grep a term, open a file, follow an import, back out, try again. It is rebuilding a picture of a codebase it mapped an hour ago and threw away. That rediscovery burns most of a run's tool calls, tokens, and latency, and it is pure overhead:

• Repeated. Every task pays the exploration cost again, from zero.
• Discarded. Whatever the agent figured out dies with the session.
• Unshared. The next teammate, and their agent, start from scratch too.

Humans onboard to a codebase once. Agents onboard every single time.

---

What Graft does

Graft builds that understanding once and writes it into your repo as a folder of linked markdown files, one node per system, API, or concept.

• Real explanations, not a list of symbols. Each node says, in plain English, what a part of the system does and how it connects to the rest, the way a senior engineer would explain it. That is the part an agent actually needs so it can skip the exploration. It is not a dump of function names.
• A real graph you can read. No embeddings, no similarity search, no index to keep warm. The graph is a set of linked files your agent opens, greps, and follows, exactly the way it reads any other file in the repo.
• Grafted into git. The graph is just files in graft/. Commit it, and anyone who clones the repo has it. No database, no server, no setup. Git does the syncing, and a stale graph shows up as a diff in review instead of rotting in some external store.
• The diff lives with the code. When a change moves things around, you see it in the graph diff in the same pull request, right next to the code that caused it.
• Your provider, your key, your model. Summaries are written by any provider you choose — OpenAI, Anthropic (native), OpenRouter, Fireworks, Groq, a LiteLLM proxy, or a local model — under your own key. The structural code graph (graft build, graft check) is deterministic tree-sitter and never calls a model at all.

---

Benchmark

An agent that reads the graph should be cheaper and faster without getting more answers wrong. That's the whole claim, so we measured it instead of asserting it.

The harness ran three variants of the same Claude Sonnet 5 agent with the same file tools: cold (explores from zero), Graft (a graft ask --source bundle pushed up front), and pull (graftfindcode/graftfileapi tools, nothing injected — context paid for only when asked). An Opus 4.8 judge scored correctness with a required-keyword floor, so a fast-but-wrong answer couldn't win by being fast. Cost is cache-aware: reads ≈0.1×, writes 1.25×, the billing model agents actually run under.

162 runs, two repos (graft itself and a real Node/Express auth service), 3 trials each, tasks split between single-file and multi-file questions.

| Metric (mean/task) | Cold Claude Code | Claude Code with graft | |---|---|---| | Cost savings ($) | 0.0429 | 0.0292 (+32%) | | Token savings | 8,070 | 4,650 (+42%) | | Tool-call savings | 4.2 | 2.3 (+46%) | | Latency savings (s) | 39.8 | 15.8 (+60%) | | Correctness | 93% | 93% (equal) |

Graft never answered worse than cold, on any corpus. The pull variant gave up most of that speed for something bigger: correctness jumped to 98%, +5 points over cold, the strongest single result in the sweep. Push when speed is what you need; pull when being right matters more.

---

SWE-bench Verified

The sweep above is our harness measuring our mechanism. So we ran the industry-standard one too — SWE-bench Verified, real GitHub issues from real repos, graded by the official swebench harness. No judge model, no similarity score: your patch is applied, the maintainers' own tests are run, and you either flip the failing test without breaking the passing ones or you don't.

50 instances, same model on both arms — Claude Sonnet 5 — same Docker images, same turn limits. The only difference is whether graft is wired in.

| Correctness & efficiency | Cold Claude Code | Claude Code with graft | Improvement | |---|---|---|---| | Correctness | 27 / 50 (54%) | 33 / 50 (66%) | +12 pts | | Token savings | 142.0M | 109.4M | +23% | | Cost savings | $52.34 | $42.43 | +19% | | Tool-call savings | 1,370 | 1,031 | +25% | | API-request savings | 2,455 | 1,875 | +24% | | Wall-clock savings | 13,094s | 8,922s | +32% |

graft resolved 33 of 50 instances against Cold Claude Code's 27 — and got there with 25% fewer tool calls, 23% fewer tokens, and 32% less wall-clock time. Every correctness win has the same shape: the baseline patches one file and misses its siblings. On django-11532 it patched 1 of the 5 files the fix requires and broke 18 previously-passing tests, twice over. On django-16263 it patched 1 of 4 and scored 102 / 103. graft found the rest — and on django-16263 did it in half the tokens and half the time.

Two harnesses, two claims: the controlled sweep says graft is cheaper and faster, SWE-bench says it's also more correct.

Correctness over all instances; tokens, cost and calls over the instances both arms resolved, for a like-for-like comparison. Official SWE-bench Verified images and official swebench 4.1.0 grader, native x86_64.

---

How the graph gets built

Graft builds the graph in two passes, both powered by a language model:

1. Read each file. Every source file is summarized once into a short description of what it does.
2. Group into nodes. Those summaries are grouped into a curated set of nodes (subsystems, key files, and concepts) with typed links between them. Graft chooses the right level of detail for you instead of making one node per file, so a big repo becomes a few dozen readable nodes.

flowchart LR
    S[Source files] --> T["Tier 1 — tree-sitterno model, no key"]
    S --> P1["Pass 1 — LLM summarizeseach file (--deep)"]
    T --> W["graft/.graph/wiring.jsonper-symbol code graph"]
    P1 --> P2["Pass 2 — group into nodes+ typed links"]
    P2 --> N["graft/*.mdmarkdown node graph"]

Every pass is cached by content hash — the LLM ones and the tree-sitter parse alike. Re-running only touches the files that changed, so the second build is fast and cheap (on this repo, 124 files: 0.74s cold, 0.18s after one edited file, 0.18s with nothing changed). graft build --no-reuse forces a cold re-parse.

That cheapness is what lets every query refresh the graph before it answers. A retrieval call stats the tree against the last build's fingerprint (~3ms), and rebuilds only if something moved — so ask/grep/callers/skeleton/map describe the code as it is right now, including edits that are unsaved to git: uncommitted, unstaged, or staged all look the same to graft, which never reads git at all. The refresh is structural and $0; it never calls the LLM. Turn it off per-command with --no-refresh, or everywhere with GRAFT_NO_REFRESH=1.

Alongside the markdown graph, graft build builds graft/.graph/wiring.json — a per-symbol code graph — plus a per-file wiring card mirroring your source tree. Tier 1 is pure tree-sitter (every function, class, and call edge; deterministic, no model, no network), which is why plain graft build needs no key. The --deep pass adds a one-line summary and a crux excerpt per symbol, cached by body hash.

---

Supported languages

Graft parses with tree-sitter at two levels of fidelity, plus an optional compiler-grade layer — all $0 and deterministic (no model, no key):

• Full-fidelity — hand-written extractors with scope-aware, cross-file call

and import resolution: TypeScript / JavaScript (incl. JSX & TSX), Python, Go, Java.

• Broad — symbols (functions, classes, methods, types, …) plus name-resolved

call edges via a generic tree-sitter extractor, one grammar per language: Rust, C, C++, C#, Ruby, PHP, Kotlin, Scala, Swift, Elixir, Solidity, OCaml, Zig, Dart.

• Compiler-grade edges (opt-in) — graft build --lsp adds precise

lsp_resolved call edges (member calls the static pass can't type) when a language server is on your PATH: rust-analyzer (Rust), clangd (C/C++), gopls (Go), pyright (Python), typescript-language-server (TS/JS). It's best-effort — with no server installed the graph is unchanged.

Twenty languages in total. A file whose language isn't listed is skipped, not indexed. Adding a broad-tier language is a small contribution — see [CREDITS.md](CREDITS.md) for the folks who added the current set.

---

What's in a node

A node is a single markdown file. Most code maps stop at an address: this thing lives in that file, on that line. That tells an agent where to look, not what it will find, so it still has to open the source and read. A Graft node holds the meaning inline, so the agent learns what it needs up front and opens the file only when it wants more.

Each node holds:

| Part | What it holds | |---|---| | Summary | A plain-English explanation of what the code does, written by the model and cached. It is there whether or not the code was ever documented, and it is regenerated when the source changes. | | Crux | The handful of lines that actually carry the logic: the guard, the skip condition, the state change. Lifted straight from the source and stored inline, so the agent sees how it works, not just what. | | Sources | The exact files the node is built from, each tracked by a content hash, so Graft can tell precisely when a node has gone stale. | | Links | Typed connections to other nodes (depends_on, part_of, uses, implements, produces), written as [[wikilinks]] your agent can follow. | | Notes | Anything you write below the generated block. It is preserved across regenerations, so your own context is never overwritten. |

That is three depths in one file: the summary says what the code does, the crux shows how, and the sources point to the rest if the agent needs it. A plain index makes it read a whole file to learn one thing. A Graft node hands it the answer inline, and the follow-up read often never happens.

The crux is stored as the code itself, not as a line range, on purpose. Line numbers drift whenever unrelated code above them shifts, but the lines that matter do not. Keeping the text, not the numbers, means the crux stays correct even as the file around it moves.

Summary, sources, links, and notes ship today in markdown nodes. The crux ships per-symbol in the code graph (graft build --deep); inlining it into markdown nodes is next.

---

What runs where

• On your machine, no key, no network: the structural code graph. graft build (wiring graph + per-file cards), graft check, and graft ask are deterministic tree-sitter — they never call a model.
• Through your provider key: the LLM-written parts — graft build --deep adds the concept nodes (file summaries + node synthesis) and the per-symbol summaries and cruxes. graft is vendor-neutral: set GRAFT_PROVIDER (openai for any OpenAI-compatible endpoint, or anthropic for the native API), your GRAFT_API_KEY, GRAFT_MODEL, and — for the openai wire format — GRAFT_BASE_URL to point at OpenRouter, Fireworks, Groq, a LiteLLM proxy, a local server, or OpenAI itself. Or pass --provider/--model/--api-key/--base-url on the command line. (OPENROUTER_API_KEY still works as a deprecated fallback.)
• No telemetry and no analytics — the only network calls are the LLM requests you configured.

See [.env.example](.env.example) for the full list of settings (model, base URL, graph directory).

---

Agent integration

One command wires Graft into the coding agents you use:

npx @nanonets/graft init
# detects your agents and writes each one's native instruction file;
# Claude Code additionally gets the live statusline + hooks below

On a terminal, init shows you every agent it knows about — flagging the ones it detected (via their config directories) and listing the exact files each would write — and wires only the ones you select. Claude Code is pre-selected; nothing else is. Selected agents get a marker-fenced Graft section in their shared instruction file — AGENTS.md (Codex, OpenCode and other CLIs that read it), GEMINI.md, .github/copilot-instructions.md — or a wholly-owned rule/skill file for the agents that use one: .claude/skills/graft/SKILL.md, .cursor/rules/graft.mdc, .kiro/steering/graft.md, .windsurf/rules/graft.md, .adal/skills/graft/SKILL.md for AdaL. Claude Code is in the second group: init writes its own skill file and never touches your CLAUDE.md. Re-running only updates Graft's own section (or replaces the owned file) and never touches the rest of your content.

With no TTY to prompt on — CI, a Dockerfile, a piped shell — init writes nothing and prints the command to run instead. Pas

…

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: NanoNets
• Source: NanoNets/Graft
• License: MIT
• Homepage: https://graft.nanonets.ai

Install and usage instructions live in the source repository linked above.