trailhq/Graft

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

What it solves

Coding agents often start every task "blind," spending significant tokens, time, and tool calls re-exploring a codebase to understand its structure and logic. Graft provides a persistent, regenerable context layer that allows agents to onboard to a codebase once rather than every single session.

How it works

Graft builds a graph of the codebase as a set of linked markdown files stored locally in a graft/ folder. It uses a two-pass process:

  1. Structural Analysis: Uses tree-sitter to create a deterministic code graph of symbols and call edges without needing an LLM.
  2. LLM Summarization: An LLM summarizes each file and groups these summaries into curated nodes (subsystems, concepts, or key files) with typed links (e.g., depends_on, implements).

This graph is stored as a local cache (ignored by git) and is automatically refreshed before queries to ensure it reflects the current working tree, including uncommitted edits.

Who it’s for

Developers using AI coding agents such as Claude Code, Cursor, Gemini, or Copilot who want to reduce latency, lower API costs, and improve the correctness of agent-generated patches in large codebases.

Highlights

  • Efficiency Gains: Benchmarks show up to 46% fewer tool calls, 42% fewer tokens, and 60% less latency compared to "cold" agent sessions.
  • Improved Correctness: Increased resolution rate on SWE-bench Verified from 54% to 66%.
  • Vendor Neutral: Supports any LLM provider via API keys (OpenAI, Anthropic, OpenRouter, etc.) or local models.
  • Broad Language Support: Supports 22 languages, with full-fidelity extractors for TypeScript/JavaScript, Python, Go, Java, and R.
  • ** uma Seamless Integration:** One-command setup (graft init) that wires the tool into the native instruction files of popular coding agents.

Related

  • Dispatch
    Graft: Open‑Source Context Layer Cuts Claude Code Token Usage by 42% and Boosts SWE‑bench Accuracy

    Graft, an open‑source code‑graph tool, reduces Claude Code token usage by 42%, cuts tool calls by 46%, and improves SWE‑bench correctness from 54% to 66% while making agents up to 4× cheaper and 3× faster.

  • Project
    vercel-labs/lat.md

    A knowledge graph system for codebases that uses interconnected markdown files and source code annotations to provide AI agents and humans with scalable, searchable context.

  • Project
    colbymchenry/codegraph

    A semantic code intelligence platform that builds a local knowledge graph of a codebase to provide AI agents with surgical context, reducing token costs and tool calls.

  • Project
    kunal12203/GrapeRoot

    A context engine for AI coding assistants that builds a semantic graph of a codebase to pre-load relevant code into prompts, reducing token costs and exploration turns.

  • Project
    Cranot/roam-code

    A local codebase intelligence layer that provides AI coding agents with pre-resolved graph facts and structural verification to reduce token usage and ensure code quality.