Graft Makes Your AI Coding Assistant Faster and Cheaper
Give AI coding assistants a reusable map of your code, reducing repeated exploration and helping them make better changes.
If you’ve watched an assistant open file after file before making a change, you’ve seen the overhead Graft is designed to cut.
It builds a reusable map of your code and how the pieces connect. Your assistant can find the relevant code, trace what depends on it, and get on with the job with less searching.
How that changes the workflow
Without a map, an agent often starts each task by grepping across the repository, opening a stack of promising files, then repeating the search when it discovers a caller, sibling implementation, or configuration edge case. Those scans are useful, but they repeatedly send broad, mostly irrelevant context into the conversation.
Graft turns that exploration into targeted questions. Instead of asking which files mention a name, the agent can jump to the defining code, trace callers and dependencies, or inspect the smallest useful slice of a file. It spends fewer tokens reconstructing the same codebase map and has a clearer picture of the change's blast radius before it edits. The result is more focused code changes—not a substitute for tests and review, but less wandering before the agent reaches the code that actually matters.
In the Graft team’s 50-task SWE-bench Verified test, the same model solved 33 issues with Graft versus 27 without it. These were real GitHub issues, checked against the projects’ own tests. The team also reported 32% less time and 19% lower cost, with cost savings measured on tasks both setups solved. That’s more issues fixed, with less time and money spent.
Graft works with Claude Code, Cursor, Codex, and Gemini. Setup takes one command, and basic code mapping runs locally without an API key. You can keep using your preferred assistant and give it better context about your project. Try Graft on a project you already work on.
