Burn

🔥 burn

🌐 Live demo →

Where did your AI budget actually go — and what did each feature cost?

burn is a local, zero-dependency ledger that reads the transcripts your coding
agents already leave on disk. It answers per repo, per model, per day,
per git branch: how many tokens you spent, how many dollars that cost, and how
much prompt caching saved you. No network. No account. No telemetry. Nothing
leaves your machine.

The ledger format itself is an open, conformance-tested protocol: ULP 1.0 —
spec'd at github.com/aashish254/ulp
(vendored here under ulp/ so one checkout runs both suites).
burn is ULP's reference implementation #1: dashboards and CI can build on the
protocol, not on this CLI.

burn — spend per repo, per model, with cache economics:

burn work — what each feature/branch actually cost:

*Both shots are real CLI output over a demo dataset of synthetic transcripts;
your own repo names, branches and numbers appear instead.*

Why

Every coding agent quietly writes down exactly what it did — Claude Code drops a
JSONL transcript per session with a full usage block and its git branch;
OpenCode keeps a SQLite db with real billed cost; Codex and Gemini CLI record
their own token histories. That data has been sitting on your laptop this whole
time. Nobody adds it up across agents, and nobody can tell you what
feature/payments-refactor actually cost.

burn answers the questions that gate agent adoption at scale:

- Which repo is eating my spend? Which feature/branch/PR?
- Am I paying opus prices for 100-token answers? (burn doctor)
- What will this pace cost me next month? (burn forecast)
- Is this PR over budget *before* it merges? (burn budget → exit 3)
- What does my team spend, with no backend? (burn export | merge | team)

Install

macOS or Linux with Node ≥…