Hi HN. I built this at work for fun because I could not answer a basic question: what did that last agent session actually cost me?

It does three things:

1. A statusline under the agent's input box while it works: which model is billing you, cost so far, dollars per hour, and how much of your rate-limit window is gone. Waste like a stuck Bash retry loop gets flagged on the line as it happens. 2. "npx aireceipts-cli" after a session prints an itemized receipt: every tool priced, plus a "same tokens on a cheaper model" line that is arithmetic on the same token counts, not a prediction. 3. "npx aireceipts-cli pr --post" attaches the cost of the sessions behind a PR as a comment. There is a real one live on the repo's PR #189.

How it works: it parses the local transcripts your agent already writes. Claude Code, Codex CLI, and Gemini CLI get full per-turn parsing (models, tools, cache tiers). OpenCode is per-message. Cursor is an honest degraded mode: its logs carry no per-turn usage, so you get session totals only. No account, and transcripts never leave your machine.

Two design rules I cared about:

- Every price is cited: vendor URL, date observed, and a quoted excerpt, checked in CI. A daily tripwire cross-checks the price table against an independent dataset and opens an issue when they disagree.

- Receipts are deterministic: same transcript in, byte-identical receipt out, golden-tested on every commit, including the receipts shown in the README.

One disclosure: there is anonymous telemetry, on by default. The reason it exists: agent log formats change underneath me all the time, and parse-failure signatures and error classes tell me which agent broke before anyone has to file an issue. It never sends code, prompts, paths, titles, or dollar amounts. "aireceipts --telemetry-show" prints exactly what a run would send, and AIRECEIPTS_TELEMETRY=off or DO_NOT_TRACK=1 turns it off.

Apache-2.0, pre-1.0, so output can still change between minor versions. I am dedicating time to this every week, so feedback, bug reports, and requests for agents whose logs you want parsed are all welcome.