Your coding agents already log every token they burn. vibebill reads those logs, joins them to git history, and answers: what did this commit cost and where did the money go?
npx vibebill doctor
Node ≥ 20 · runs read-only on your local agent logs · no telemetry, no cloud, one opt-in network call (--refresh-pricing)
Real output. This project was built by AI coding agents, and vibebill metered the build commit by commit. The last line is the token-conservation invariant: every parsed token lands in exactly one bucket, checked on every run.
- reads local agent transcriptsclaude-code · codex · gemini · aider
- attributes usage to commitsfiles + timing + branch, confidence-marked
- prices every call exactlybigint nano-USD, never floats
- counterfactual repricing
reprice --model deepseek-chat - markdown ledgers & git notes
report --md·notes sync - PR costing via gh
vibebill pr --comment - privacymetadata only — prompts never read
npx vibebill doctor
environment check: adapters found, parse coverage, pricing table
npx vibebill log
git log with a cost column — the screenshot above
npx vibebill summary
totals, cost by model, cache savings, waste, invariant status
Attribution is probabilistic. Every figure carries a confidence marker (● ◐ ○) and the evidence is inspectable with vibebill show.
Subscription plans show API-equivalent value, not billed dollars. Repriced numbers are same-traffic counterfactuals, not workload simulations.
Unknown models are never guessed: tokens shown, cost $—. Cursor is unsupported — its usage data lives server-side where no local tool can measure it.