Инструмент Graft повышает эффективность и точность ИИ-разработчиков на 12% благодаря контекстному анализу кода

Turbocharge Claude Code, Cursor, Codex, Gemini & every coding agent: faster, cheaper, with contextual understanding specific to your codebase. 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 ↓ · SWE-bench ↓ · Per-repo numbers ↓ - Quick start - The problem - What Graft does - Benchmark - SWE-bench Verified - How the graph gets built - Supported languages - What's in a node - What runs where - Agent integration — MCP server · Claude Code (deep integration) - CLI - Search & orient (graft grep /graft map ) - Monorepos & multi-repo folders - Visualize it (graft viz ) - Tested on your popular repos - Development - License 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. 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. 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. 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 (graft_find_code/graft_file_api 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. 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. 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. 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. Graft builds the graph in two passes, both powered by a language model: - Read each file. Every source file is summarized once into a short description of what it does. - 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-sitter<br/>no model, no key"] S --> P1["Pass 1 — LLM summarizes<br/>each file (--deep)"] T --> W["graft/.graph/wiring.json<br/>per-symbol code graph"] P1 --> P2["Pass 2 — group into nodes<br/>+ typed links"] P2 --> N["graft/*.md<br/>markdown 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. Git determines the visible file set; freshness compares the working-tree bytes rather than commit or index state. 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. 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, Clojure. - Compiler-grade edges (opt-in) — graft build --lsp adds preciselsp_resolved call edges (member calls the static pass can't type) when a language server is on yourPATH : 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-one 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 for the folks who added the current set. 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: 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. - On your machine, no key, no network: the structural code graph. graft build (wiring graph + per-file cards),graft check , andgraft 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: setGRAFT_PROVIDER (openai for any OpenAI-compatible endpoint, oranthropic for the native API), yourGRAFT_API_KEY ,GRAFT_MODEL , and — for theopenai 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 for the full list of settings (model, base URL, graph directory). 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. Pass --agents <ids> or --yes to make a scripted run explicit. Selecting the agents host also touches your user-level Codex config, when ~/.codex/ exists: Both configs are user-level, so they apply to every repo you open with Codex, not just this one. The picker labels these machine-wide, --dry-run lists them in their own section, and --no-global skips them while still wiring AGENTS.md. graft init also registers Graft's MCP server with agents that support it, so these six tools appear natively, no shell required. Claude Code gets this too: graft init writes the server into the project's .mcp.json (restart Claude Code to load it). Skip with --no-mcp; run it manually with graft mcp [dir]. Register it by hand if your agent needs it explicit: { "mcpServers": { "graft": { "command": "npx", "args": ["-y", "@nanonets/graft", "mcp"] } } } Where a CLI agent supports user-level hooks.json, init also installs Graft's post-edit hook — blast-radius warnings and automatic $0 graph re-sync after edits (skip with --no-hooks). graft init always wires up Claude Code, and Claude Code gets more than the skill file above. From then on, any Claude Code session opened in the repo gets: - a live statusline — graph size, % enriched, and a ⚠ N stale warning when the code has moved ahead of the graph - auto-sync — every graft query brings the graph up to date first, so an answer always describes the code as it is right now, uncommitted edits included. A query refreshes only what it reads; the markdown under graft/ is refreshed by the background rebuild at the end of a turn that touched code. Both are structural and$0 — auto-sync never calls the LLM on its own - context on tap — each prompt pulls the matching nodes into the session; editing a file surfaces what depends on it ("blast radius"); new sessions start with the repo map install → init → hooks keep the graph fresh every session edit a file → blast radius appears inline → graph auto-resyncs → confirmed in graft viz graft init is idempotent and never clobbers your existing .claude/settings.json — it merges its blocks and leaves the rest alone. Want the LLM summaries too? Run graft build --deep (with a key) whenever you like; auto-sync will never do it for you. graft build [dir] # build graft/ from the code at [dir]: wiring graph + per-file cards (no LLM, no key) graft build --deep # add the LLM layer: concept nodes + per-symbol summary/crux (cached) graft build --extensions .ts .py # only include these code extensions graft build --no-reuse # re-parse every file instead of replaying unchanged ones from cache graft build --follow-submodules # include initialized submodules; persist the choice for builds + MCP refresh graft build --no-follow-submodules # exclude submodules again and persist that choice (the default) graft ask "<task>" [dir] # query the graph — ranked nodes + exact file:line (no LLM, no key) graft ask "<task>" --json # machine-readable result graft ask "<task>" --in <scope> # narrow to one sub-project of a monorepo/multi-repo folder (see below) graft skeleton <file> [dir] # every signature in one file, no bodies — the API surface for ~1/10th the tokens (no LLM, no key) graft callers <symbol> [dir] # who calls/references/imports/implements/extends a symbol (no LLM, no key) graft callers <symbol> --direction out # the reverse: what the symbol itself calls/references (was `graft callees`) graft callers <symbol> -d N # walk transitively out to depth N — full blast radius (was `graft impact`) graft grep "<regex>" [dir] # exhaustive regex search over indexed files, grouped by enclosing symbol (no LLM, no key) graft grep "<regex>" --in <path> # narrow to files at or under this path prefix graft grep "<regex>" -i --fixed # case-insensitive; treat the pattern as a literal string, not a regex graft map [dir] # token-budgeted repo orientation — dir clusters, hubs, hotspots (no LLM, no key) graft map --max-dirs N # raise/lower the number of directories shown graft check [dir] # fail (exit 1) if graft/ has drifted from the code (never auto-refreshes — it's the drift report) graft check --json # print the drift report as JSON # ask / skeleton / callers / grep / map all refresh the graph first if the working tree moved: # --no-refresh # answer from the graph exactly as it is on disk # GRAFT_NO_REFRESH=1 # same, for every command # GRAFT_REFRESH=hash # hash every file instead of trusting size+mtime graft viz [dir] # see the graph: serves an interactive viewer on localhost graft viz --port 5000 --no-open # pick a port; don't auto-open the browser graft init [dir] # pick which agents to wire (prompts on a terminal; writes nothing until you choose) graft init --dry-run # list every file it would touch, then exit graft init --agents cursor kiro # wire only these agents, no prompt (ids: agents, cursor, gemini, copilot, kiro, windsurf, adal, claude) graft init --yes # no prompt; wire every detected agent graft init --no-global # skip writes outside this repo (~/.codex/ config + hooks) graft init --no-build # wire the files only; don't build the graph graft init --all-agents # wire every known agent, detected or not graft init --list-agents # list known agent ids and exit graft version # print the installed + latest published npm version graft upgrade # npm install -g the latest published version # a new version is announced automatically (checked once a day); # after upgrading, the next session refreshes this repo's wiring itself # global graft --dir <path> # use a context dir other than <repo>/graft graft --version, -v # print the installed version and exit Method calls resolve through the receiver's type — constructor assignments (self.router = APIRouter()) and type annotations, not just the call-site name — so callers/grep --in return calls bound to the right type on method-heavy code, not every method anywhere with that name. graft grep "<regex>" is exhaustive over every indexed file and groups hits by enclosing symbol, ranked by the same in-edge coupling graft map uses — built for "every occurrence of this pattern" tasks where graft ask's ranked top-N isn't enough: "NEEDLE" — 2 hits in 2 symbols across 1 files (searched 1 indexed files) heavilyCalled · function · src/a.ts:L1-L3 · 3 in-edges L2: console.log("NEEDLE hit in heavilyCalled"); rarelyCalled · function · src/a.ts:L4-L6 · 0 in-edges L5: console.log("NEEDLE hit in rarelyCalled"); graft map is a token-budgeted first look at a repo — directory clusters with file/symbol counts, each dir's local hubs, and the global hotspots — all ranked by in-degree, no LLM, no key: repo map — 113 files · 687 symbols · 2186 edges · typescript src/ 63 files · 527 symbols hubs: contextDirFor (node-file.ts, 21←), wiringPath (write.ts, 14←), buildGraph (build.ts, 11←) test/ 43 files · 102 symbols hubs: edge (graph-traverse.test.ts, 4←), graphOf (graph-traverse.test.ts, 4←), fileNode (graph-map.test.ts, 3←) viewer/ 5 files · 58 symbols hubs: $ (main.ts, 9←), activeGraph (main.ts, 5←), cvar (data.ts, 5←) scripts/ 2 files · 0 symbols hotspots: contextDirFor · function · src/context/node-file.ts:L100-L103 · 21← wiringPath · function · src/graph/write.ts:L20-L22 · 14← buildGraph · function · src/graph/build.ts:L104-L218 · 11← ... Graft supports these layouts: - A monorepo with one .git (apnpm-workspace.yaml /package.jsonworkspaces , or per-packagego.mod /pyproject.toml /Cargo.toml ) —graft build discovers each sub-project as a ranking scope.ask /map rank every scope on its own terms and fuse the results, so the biggest sub-project can't drown a small one; hits carry[scope/] labels, andgraft map groups its directory clusters by scope first. - A Git superproject with initialized submodules — submodules stay excluded by default. Run graft build --follow-submodules to fold initialized gitlinks into one graph, prefixing child paths (for example,deps/parser/src/index.ts ) while honoring each submodule's own Git ignore rules. Visible untracked files are included too; uninitialized submodules remain absent untilgit submodule update --init checks them out. The choice is saved in.graft/config.json , so later no-flag builds and MCP automatic refreshes behave the same way. Rungraft build --no-follow-submodules to restore and persist the default boundary. - A folder of separate git repos (no .git at the top) —graft build auto-splits: each child gets its own (git-ignored)graft/ , and the parent gets agraft/workspace.json index. Queries from the parent federate across every child, always labeled<child>/ . Rungraft build inside a child to work on just that repo. In every layout, narrow to one sub-project with graft ask "<task>" --in <scope>/ once you know where you're working. graft init at the parent of a multi-repo folder wires every child repo too, not just the parent — an agent session opens at a repo root and reads its instruction files from there, so each child needs its own. A session started in the parent gets the federated view; one started in a child sees that repo alone. Commands also find the graph from a subdirectory: with no [dir] argument they walk up to the nearest graft/, so graft ask works from src/deep/inside/ without a cd to the repo root. graft viz opens a local, interactive view of both graphs — no install, no dev server; the viewer ships prebuilt inside the package. search → jump to a node → dependency graph lights up - Context tab — the architecture graph from graft/*.md . Nodes colored by type, sized by connectedness. - Code tab — the per-symbol graph from graft/.graph/wiring.json (rungraft build first). - Outline tab — the file → class → method hierarchy as a collapsible tree. Edges speak the code's language. Every link is one of a closed set of verbs, each answering a question someone building or reviewing code actually asks: Select a node and its edges take on direction: amber = what it depends on, teal = what depends on it, with the verb written on each highlighted edge. Chips above the canvas filter by verb; tree-sitter-extracted edges draw solid while LLM-inferred ones draw dashed. The viewer live-reloads when graft/ changes on disk. Older graphs with vague verbs (influences, supports) are normalized on load — no regeneration needed. The benchmarks measure the mechanism. The real test is whether graft helps an agent ship real changes on code people actually run, not just answer questions. So we benchmark it on popular open-source repos: 15 tasks each, 10 real developer questions plus 5 actual implementation tasks (real merged pull requests, each re-implemented from its base commit and scored against the files the maintainers actually changed). Same agent (Claude Opus), same file tools; the only difference is whether graft is wired in. Across these repos graft runs up to 4× cheaper and 3× faster, with better or no loss of correctness: it reproduces the real merged PRs by touching the same files the maintainers did. Per-repo detail below. Cheaper and faster with no loss of correctness: graft reproduced all five merged PRs, touching the same files the maintainers did. The gap is widest on cross-file understanding — "how does auth work across OAuth2 providers" dropped from $2.19 to $0.84. The 10 questions we asked - Orientation — Give me a map of PocketBase's architecture: the main subsystems and how an HTTP request flows through to the database. - Entry-point trace — Trace end-to-end what happens when a client creates a record via the REST API, from route handler to database write. - Feature location — I want to add a brand-new collection field type. Where do I hook it in, and which pieces must change? - Bug localization — Realtime subscriptions silently stop delivering events after a while. Where would you start looking, and why? - Blast radius — If I change the signature of the record-validation logic, what depends on it and what could break? - Cross-file synthesis — How does auth work across OAuth2 providers: where are tokens issued, validated, stored, and refreshed? - Extensibility — How do I use PocketBase as a Go framework to register a custom route plus an on-record-create hook? - Security discovery — Where is user input validated, and where are collection API access rules enforced before a query runs? - Public API — As an external app, how do I authenticate and then list and filter records over the REST API? - Test verification — Where are the tests for the record CRUD API, and what do they assert about access rules? The 5 merged PRs we re-implemented Each PR was reset to its base commit; graft's diff was scored against the files the merged PR changed. Method Two clones of PocketBase at the same commit: one wired with graft init, one untouched and verified graft-free. Each task run headless (claude -p, Claude Opus) with an empty MCP config. Understanding questions were graded by whether the answer pointed to the right files and functions; PR tasks were scored on whether the agent's diff touched the same files as the merged PR. Every transcript was audited to confirm graft was actually used in the graft arm and absent from the standard arm. git clone https://github.com/NanoNets/context-graph-engine.git && cd context-graph-engine npm install npm run build npm test npm run cli -- build --deep . # run the CLI from source MIT. See LICENSE.