I started building this because I watched Claude Code read the same 7 files to answer "who calls parseToken?" for the third time in one week. At some point you either accept the wasted cost or you fix it. Code structure doesn't change between your sessions, but the agent acts like it does. So I built a local call graph that persists. Now when it needs to know who calls something, it gets the answer in under 1ms with zero file reads. I benchmarked it honestly. 61% fewer tokens overall, $8.88 to $3.68 per session. I published the results where it doesn't help too — trace and flow questions are neutral. That felt important to say upfront. A few things I'm genuinely happy about: every response includes file:line@sha so the agent cites a verifiable location, not a guess. gp_impact with since: "main" scopes blast radius to only your branch's changed files, not the entire codebase history. And if you use git worktrees, two branches get two separate indexes automatically. Ask me anything about the benchmark, the tree-sitter implementation, or how it compares to other MCP code tools. What structural questions does your agent ask most? I want the next benchmark to cover real usage, not just the questions I thought mattered.

GraphPilot

Stop paying for files your agent already read.

Stop paying for files your agent already read.