agent-ledger v0.1.1What did that task cost?
Claude Code already writes every token it spends to a transcript on your disk. agent-ledger reads them and tells you, per session, branch or day, what each piece of work cost — and why.
npx @allan_nava/agent-ledgerNode 18+. No dependencies, no API key, nothing leaves the machine.
One row per ticket
branch sessions resp peak ctx billable cache read hit cost ----------------- -------- ---- ------------- --------- ---------- ----- ------ feat/rate-limits 3 180 252,446 (25%) 1,053,962 15,623,949 97% $9.30 fix/webhook-retry 1 26 106,379 (11%) 115,987 1,523,596 99.9% $0.918 chore/deps 1 18 79,010 (8%) 80,013 824,489 99.9% $0.578 refactor/billing 1 140 546,356 (55%) 677,007 39,242,202 99.9% $15.10 ----------------- -------- ---- ------------- --------- ---------- ----- ------ total 6 364 1,926,969 57,214,236 99% $25.89
Sample data, generated by scripts/sample.mjs — the shape of a real week, not a real project. --by branch puts every session spent on a branch in one row. Here the rate limiter, done in three short sessions, never passes 25% of the window; the one long billing session reaches 55% and costs more than the whole feature. The same data as the --html report →
What it gets right
Each response once
Claude Code writes one API response as several transcript entries, each repeating its usage. Summing entries roughly doubles every figure; agent-ledger counts by message ID.
Cache priced by TTL
Cache writes are 1.25× for five minutes and 2× for an hour; reads are 0.025× to 0.1× depending on the model. The transcript says which — so the cost does too.
Subagents included
A subagent's transcript sits beside its parent's. Its tokens land in the session that launched it, so a cheap-looking session that fanned out shows its real cost.
Unknown is not free
A model with no price in the dated table is marked and left out of the total — never silently priced at zero.
On a subscription
Pro and Max are not billed per token. There the cost column is what the same traffic would cost on the API: a way to compare tasks, models and habits, not your bill.