AI coding tools re-read your entire codebase every turn. GrapeRoot fixes this. It builds an AST + dependency graph of your project locally, then injects only the relevant files (~4,000 tokens) before the AI starts thinking. No cloud, no tool-call overhead, no code leaves your machine. 60–70% token savings, measurable on real codebases. Works with Claude Code, Cursor, Codex, Gemini CLI, and GitHub Copilot. Open source, zero config, 150ms injection.

Hey PH! I built GrapeRoot after getting frustrated watching Claude Code re-read the same file on every single turn of a conversation. The fix wasn't another chat wrapper, it was fixing context handling upstream. GrapeRoot builds an AST + dependency graph of your codebase and injects only the files relevant to your current query, before the model even starts thinking. On a real 2000+ file production codebase: 84.7/100 quality score,60% cost reduction on complex tasks. Everything runs locally, your code never leaves your machine. Works as an MCP server with Claude Code, Cursor, Codex, Gemini CLI, Copilot. Would love to hear how much your Claude Code bill is eating through your budget and whether this helps. Ask me anything 🙌

Re-reading the whole repo every session is one of those hidden Ai coding costs people notice only after usage grows. Does GrapeRoot work best for large monorepos, or do smaller projects also see meaningful token savings?

GrapeRoot

Cut your AI Bill by 60–70% with a local codebase graph!

Cut your AI Bill by 60–70% with a local codebase graph!