ashthejuan

TokenPilot is a local-first AI engineering cost intelligence tool that sits between coding agents and the model API, records every request with project/session/task labels, and measures the cost of successfully completed work, not just tokens burned. It runs seven deterministic detectors to surface waste (retries, repeated context, runaway sessions, and more), uses Meta Muse to investigate findings from an evidence bundle and propose concrete, testable changes, then proves those optimizations with paired benchmarks that only label **Verified savings** when the alternative still passes acceptance tests. The result is a clear loop—observe, attribute, detect, explain, experiment, verify, report—so developers can see where AI coding spend goes, why it’s wasteful, and how much a change actually saves without sacrificing correctness.