GitHub - anandgupta42/aireceipts: Your AI coding agent just billed you. Here's the receipt. Cost receipts for Claude Code, Codex, Cursor, Gemini & opencode — no account, no upload.

Your AI coding agent used millions of tokens. Here is the observable cost floor.

_{not a mockup — a real receipt on a merged PR of this repo. The screenshot predates the current ≥ $X notation. Read it live.}

The meter runs while the agent drives · the receipt prints when the ride ends · and sticks with the PR

Why this exists. AI coding-agent usage is usually invisible — you see the diff, not the token ledger. aireceipts reconstructs an observable Standard-API list-price floor: live in your status bar, itemized when the session ends, attached to the PR. It is not an invoice or subscription allocation. Local and deterministic — transcripts never leave your machine, and a shared receipt carries figures plus the session's one-line title (drawn from your opening prompt, truncated) — never your diff, file contents, or transcript (how).

Start here — the meter, the receipt, the PR

Try it in ten seconds: npx aireceipts-cli — no install, no account (--demo shows a bundled example if you have no sessions yet). Then let it run like a cab ride:

1 · While the agent works — the meter. One settings line (setup) pins aireceipts statusline under Claude Code's input box — and tmux, starship, or PowerShell give Codex and OpenCode the same bar — ticking up as the session runs:

The line, segment by segment:

[aireceipts] Opus · ≥$4.20 · ≥$9/hr · 128k · ctx 42% · 5h 24% ↺2h13m
             │      │       │       │      │         └ how much of your 5-hour cap is gone · when it resets
             │      │       │       │      └ how full the context window is
             │      │       │       └ how many tokens the session has used
             │      │       └ observable Standard-API floor per hour
             │      └ observable floor so far — cited prices, subagents included
             └ which model is active right now

When a detector finds a stuck retry loop, a heuristic pattern flag appears right on the line; it is evidence to inspect, not a proven savings claim.

2 · When the session ends — the receipt. npx aireceipts-cli prints the itemized receipt — cited floors on priceable tool rows, tokens-only buckets where evidence is insufficient, heuristic patterns flagged (loops, context thrash, trivial spans), and the cheaper-model line as arithmetic, not a prediction. The exact bytes:

- - - - - - - - - - - - - - - - - - - - - - - - -
                    AIRECEIPTS                    
 “Add email format validation to the signup for…” 
 Claude Code · Jun 18 2026 09:30:30 UTC · 10m 30s 
    claude-opus-4-8 87% · claude-sonnet-5 13%     
         cache served 85% of input tokens         

pre-edit: 11% of priced floor (1/10 turns)
  (share before the first named edit tool)

Bash..........................≥ $0.0517  (3 calls)
Edit..........................≥ $0.0455  (2 calls)
(thinking/reply)..............≥ $0.0310  (2 turns)
Write.........................≥ $0.0290  (2 calls)
Read...........................≥ $0.0192  (1 call)
--------------------------------------------------
TOTAL....................................≥ $0.1764
standard API-equivalent floor; not an invoice
same tokens on claude-haiku-4-5..........≥ $0.0392
  (78% lower observable floor)
  (arithmetic, not a prediction)
- - - - - - - - - - - - - - - - - - - - - - - - -
                npx aireceipts-cli                
         github.com/anandgupta42/receipts         
- - - - - - - - - - - - - - - - - - - - - - - - -

The final line identifies the open-source project that generated the receipt; the line above it is the install command. PR comments and their HTML artifacts make the same source-and-install destination clickable.

Every human ≥ $X is rounded down. Additive spend rows share one adaptive precision — two decimals for exact cents, normally four for fractional cents, and up to twelve for tiny positive evidence — and visibly sum to a TOTAL that never exceeds the raw machine aggregate. No row is rounded upward; if floating-point addition serializes just below the row-unit sum, the largest row is lowered by the excess unit(s). --json/--csv retain raw precision and explicit lower-bound semantics.

_{pre-edit is the share of the observable priced floor before the first edit-tool call (reading a receipt).}

3 · When the PR ships — the receipt sticks with it. npx aireceipts-cli pr --post attaches the observable cost floor of the sessions behind a PR as a comment. Generation stays local; a drop-in CI check can require every PR to carry one — a real one, live: PR #189. Want it automatic? A paste-ready pre-push hook attaches the receipt ref whenever a local session matches the pushed branch, from any agent, and the CI check posts the comment for you.

What it tells you

• Which model is active, and the observable Standard-API floor so far — on every prompt, before the session is over, with your 5-hour cap countdown next to it.
• Where the observable usage went — priceable rows carry cited ≥ floors; unmatched or unattributed components stay as tokens. Heuristic flags identify stuck loops, context thrash, and trivial turns without claiming avoidability.
• The observable floor behind a PR — attributed across the sessions that built it, with same-tokens-on-a-cheaper-model arithmetic.

Supported agents

Model prices move. A daily advisory tripwire cross-checks data/prices/ against an independent dataset and opens an issue when they disagree; every table change lands as a cited price-table PR.

Install — or let your agent do it

No account, no API key. The command is then aireceipts; or skip the install and prefix any command as npx aireceipts-cli. Full walkthrough: getting started.

Easiest: let your agent do it

You already have an AI coding agent. In Claude Code, Codex, or any agent with shell access, paste:

Read https://raw.githubusercontent.com/anandgupta42/receipts/main/README.md
and follow the "Instructions for agents" section: install aireceipts globally
and set up the Claude Code statusline.

The agent handles the npm install and the settings edit, and shows you a receipt when it's done.

Instructions for agents

If you are an AI agent asked to install aireceipts, do the following:

1. Install the CLI globally. Run npm install -g aireceipts-cli, then confirm with aireceipts --version. If global installs are not permitted on this machine, fall back to npx aireceipts-cli in the steps below and tell the user why.

2. Set up the Claude Code statusline (skip if Claude Code is not installed, and say so). Add a statusLine entry to ~/.claude/settings.json:

   {
     "statusLine": {
       "type": "command",
       "command": "aireceipts statusline"
     }
   }

   Merge into the existing JSON; do not overwrite other settings. If a statusLine entry already exists, show it to the user and ask before replacing it. If aireceipts is not on the PATH that GUI-launched apps see, use the absolute path from which aireceipts as the command value.

3. Verify. Run aireceipts to print a receipt for the newest session (aireceipts --demo if the user has none yet) and aireceipts setup for detected agents. The statusline appears on the next Claude Code prompt.

4. Show the user the result. Print the receipt and tell them the statusline is active. Optional next steps: aireceipts install-hook (consent-gated end-of-session mini-receipt) and aireceipts integrations (snippets for Codex, OpenCode, Cursor, and GitHub PR receipts).

Everything else it does

The honesty rules

Every displayed dollar uses a cited row (vendor URL, observed date, excerpt — checked by CI). Receipts are deterministic and golden-tested. No matching model row means tokens-only; cached reads/writes without a cited applicable rate contribute zero with a caveat. Codex supports deterministic request-level Standard-API floor arithmetic, not an exact invoice: traces omit cache writes, billing/auth route, request/invoice ids, discounts, and credits. Comparisons re-price identical tokens; they never predict. What a receipt proves, and what it can't: docs/trust.md · aireceipts --methodology.

Telemetry

Anonymous diagnostics and usage signals, on by default (in CI too) — error classes, duration buckets, parse-failure signatures, feature enums, and coarse buckets. Never code, prompts, paths, titles, or dollar amounts. See exactly what a run would send: aireceipts --telemetry-show. Kill it: AIRECEIPTS_TELEMETRY=off or DO_NOT_TRACK=1. Schema and rationale: docs/telemetry.md.

Docs

User guide — get started, every command, pricing, troubleshooting (hosted docs · site). Also: FAQ · What a receipt proves · PR receipts · JSON schema · statusline.

Versioning & contributing

Pre-1.0 (0.x): minor versions may change behavior or output, patch versions are fixes only. The receipt's byte-stability contract is the compatibility surface — a change that breaks it is a major bump (changelog · releases). aireceipts is designed and largely built by AI agents under a spec-driven harness — adversarially validated specs, mutation-tested money paths, byte-golden outputs, and PRs that carry the receipt of the session that built them (how and why). Human PRs are welcome and run the same gates: CONTRIBUTING.md.

License

Apache-2.0.

Buy me a samosa

Every open-source project asks you to buy the maintainer a coffee. Not this one — buy me a samosa, and I'll explain.