 | TokenDiet: code graph for AI coding agentsCut the tokens Claude Code, Codex, Cursor and Copilot spend exploring your repo. One click builds a local code graph, wires up the agents, and carries a repo map + notes into every new session. Shows how much it saved. Installation Launch VS Code Quick Open ( Ctrl+P), paste the following command, and press enter.Copied to clipboard |
TokenDietYour AI coding agent spends most of its tokens re-reading your repo and carrying it around in context.
TokenDiet builds a local code graph of the workspace and gives Claude Code, Codex, Cursor and Copilot five
small tools to ask it instead: where is X, who uses it, what breaks if I change this file. It also
carries a repo map and the agent's own notes into every new session, so Use
First enable sets up a private Python environment (needs Python 3.10+ on the machine) — about a minute, once. What Enable writes in your repo
TokenDiet: Disable removes exactly these additions and leaves the rest of each file untouched. The numbersThe report shows, from the tool's own call log: tokens served vs what a careful agent (grep + ranged reads) and a typical one (grep + whole-file reads) would have pulled in, plus Claude Code session medians before vs after enabling. On 4 open-source repos the tools used 55–81% fewer tokens than the careful baseline per navigation task, with 95–100% correct definitions and 0.70–0.97 usage recall. Estimates under-count by design. Everything runs locally. No code leaves your machine. License and sourceMIT © 2026 Genga28. Source, issues and benchmarks: https://github.com/Genga28/TokenDiet |