Graft

Graft builds a context layer for a codebase so a coding agent does not re-explore it on every task. graft build parses the repository with tree-sitter into a per-symbol wiring graph and per-file cards; graft build --deep adds a language-model layer that summarises each file and groups those summaries into a curated set of markdown nodes — one per subsystem, API, or concept — with typed links between them. The graph is written to graft/ and added to .gitignore as a regenerable local cache; what a team commits is the wiring graft init drops in.

A node is one markdown file. It holds a plain-English summary of what that part of the system does, a crux of the few lines that carry the logic — stored as the code itself rather than as line numbers, so it survives unrelated edits above it — the content-hashed source files it was built from, typed [[wikilink]] connections such as depends_on and produces, and any notes you write below the generated block, which survive regeneration.

Every query refreshes the graph against the working tree first: a structural check of roughly 3 ms that never calls a model, so ask, grep, callers, skeleton, and map describe the code as it is right now, uncommitted edits included. Only the --deep layer calls a model, through whichever provider, key, model, and base URL you configure.

Two measurements compare the same agent with and without Graft: a 162-run sweep across two repositories, using a Claude Sonnet 5 agent scored by a separate Opus 4.8 judge against a required-keyword floor, and 50 SWE-bench Verified instances on the same model, images, and turn limits, graded by the official harness.

Measure	Cold Claude Code	Claude Code with Graft
Tool calls per task (sweep)	4.2	2.3
Tokens per task (sweep)	8,070	4,650
Latency per task (sweep)	39.8 s	15.8 s
Correctness (sweep)	93%	93%
SWE-bench Verified resolved	27 / 50	33 / 50

graft init detects the coding agents on the machine and writes each one's native instruction or rule file — a marker-fenced section in AGENTS.md, GEMINI.md, or .github/copilot-instructions.md, or an owned skill file for Claude Code, Cursor, Kiro, Windsurf, Grok, and AdaL — then registers a Model Context Protocol server exposing six tools.

Features

• Two-tier graph: a deterministic tree-sitter wiring graph plus an optional model-written concept layer
• Markdown nodes: summary, crux, content-hashed sources, typed wikilinks, and preserved user notes per node
• Fresh on every query: a structural restat rebuilds only what moved, so answers cover uncommitted edits
• CLI: ask, grep, callers, skeleton, and map for search, blast radius, file API surfaces, and repository orientation
• MCP tools: graftfindcode, graftfileapi, grafttracecalls, graftfindall, graftrepomap, and graftcheckfreshness
• Agent wiring: graft init writes instruction or skill files for Claude Code, Codex, Cursor, Gemini, Copilot, Kiro, Windsurf, Grok, AdaL, and other CLIs, with --dry-run and --agents for control
• Claude Code hooks: a statusline with graph size and staleness, background re-sync after each turn, and blast-radius warnings when a file is edited
• Language coverage: 23 languages — scope-aware extractors for TypeScript/JavaScript, Python, Go, Java, Kotlin, PHP, Swift, and R, plus broad symbol and call extraction for Rust, C, C++, C#, Ruby, Scala, Elixir, Solidity, and others
• Optional compiler-grade edges: graft build --lsp adds resolved call edges via rust-analyzer, clangd, gopls, pyright, or typescript-language-server
• Provider choice: the --deep layer runs against OpenAI, the native Anthropic API, OpenRouter, Fireworks, Groq, OrcaRouter, a LiteLLM proxy, or a local model
• Content-hash caching: every pass replays unchanged files, so a rebuild after one edit is a fraction of a cold build