Graft

Graft is a context layer for large codebases. It builds an understanding of a repository once and writes it as a folder of linked markdown files, one node per subsystem, API or concept, so that Claude Code, Cursor, Codex, Gemini and other coding agents start a task oriented instead of re-exploring the repository with grep and file reads. The graph is a set of plain files the agent opens, greps and follows like any other file: no embeddings, no similarity search, no database and no server. The generated graph is a regenerable local cache, and only the small agent wiring is committed.

Each node holds a plain-English summary of what the code does, a crux (the handful of source lines that carry the logic, stored as text rather than line numbers so it stays correct as the file shifts), the source files it was built from tracked by content hash, typed wikilinks to other nodes (depends_on, part_of, uses, implements, produces), and a notes section that is preserved across regenerations. Alongside the markdown graph, a per-symbol code graph records every function, class and call edge, with a per-file wiring card mirroring the source tree.

Building runs in tiers. The structural code graph is deterministic tree-sitter that needs no model or key and runs locally, and covers 23 languages, nine of them with full-fidelity extractors (scope-aware, cross-file call and import resolution) and the rest with a broader symbol-and-call tier. An opt-in pass adds compiler-grade call edges from a language server when one is installed. The deep pass uses a model of your choice (OpenAI, Anthropic, OpenRouter, Fireworks, Groq, OrcaRouter, a LiteLLM proxy, or a local model) to summarize each file and group the summaries into nodes with typed links. Every pass is cached by content hash, and every query refreshes the graph against the working tree first, so answers describe the code as it is now, uncommitted edits included.

One init command detects the coding agents in use, writes each one's native instruction file, registers an MCP server with six tools (ranked code search with source inlined, a file's API surface without bodies, call tracing in either direction for blast radius, regex search grouped by enclosing symbol, a repo map, and a freshness check), and installs a post-edit hook that prints blast-radius warnings and re-syncs the graph. Claude Code additionally gets a live statusline, matching nodes pulled into each prompt, and the repo map at session start. The CLI covers ask, skeleton, callers, grep, map and blast, a viewer shows the architecture graph, the per-symbol graph and a file outline with edges labeled by a closed set of verbs, and monorepos, submodules and multi-repo folders are supported. The only network calls are the configured model requests, a daily version check and an anonymous, opt-out usage ping.

The project reports two benchmarks: a controlled sweep (162 runs over two repositories, Claude Sonnet 5 on both arms, an Opus 4.8 judge with a required-keyword floor) and SWE-bench Verified (50 instances, same model and turn limits, graded by the official harness).

Metric	Cold Claude Code	With Graft
Sweep: cost per task	$0.0429	$0.0292
Sweep: tokens per task	8,070	4,650
Sweep: tool calls per task	4.2	2.3
Sweep: latency per task	39.8 s	15.8 s
Sweep: correctness	93%	93%
SWE-bench Verified: resolved	27/50 (54%)	33/50 (66%)
SWE-bench Verified: tokens	142.0M	109.4M
SWE-bench Verified: wall clock	13,094 s	8,922 s