AgentReceipt

For Claude Code and Codex CLI

What did that PR actually cost?

Drop in a coding-agent session and get an itemized receipt: the cost of every step, and why the expensive ones were expensive, with what fixing each would have saved.

Drop a session log here

Claude Code ~/.claude/projects/…/*.jsonl, a Codex rollout ~/.codex/sessions/…/rollout-*.jsonl, or the claude-code-action execution file.

Parsed in your browser. Your prompts and code never leave this tab.

Or try the .

Not just a total. A reason.

Token trackers tell you a session cost $38. AgentReceipt tells you $21 of it was one cache miss after lunch and a test loop you could have stopped at attempt two.

Idle cache miss

You stepped away for more than 5 minutes, the prompt cache expired, and the whole context was written again at 12.5x the cached price.

Cache rewrite

An MCP server, CLAUDE.md edit or model switch mid-task invalidated the cache with no idle gap at all.

Retry loop

The same file edited, or the tests run, again and again after failures. We count the loops and price the turns after the second.

Oversized tool output

A 40k-token log dump enters the context once and is paid for again as cache reads in every turn after it.

Model overkill

Top-tier model turns that only read, grep or list files, with the estimated difference on a mid-tier model.

Compaction

The session hit its context limit and paid to summarize and rebuild the context.

A receipt for every finished task

01

Try it in the browser

Drop any session. No account, nothing to install, nothing uploaded.

02

Add the one-line hook

A Claude Code Stop hook (or a Codex wrapper) sends each finished session’s numbers to your workspace.

03

Get it on the PR

The GitHub Action step posts the receipt as a PR comment, so cost per PR is visible where the work is reviewed.

How it compares

Good tools exist for request-level tracing and daily totals. AgentReceipt is about the unit your manager or client cares about: the finished task.

 AgentReceiptccusageLangfuse / HeliconeAgentMeter
Cost per PR or ticketYesNo (daily/session)With custom taggingYes
Explains why a step was expensiveYes, with savingsNoNoNo
Try without installingYes, in the browserCLI installSDK or proxyGitHub Action
Share link and PR commentYesNoNoPR comment
Prompts stay on your machineYesYesNo, tracedVaries

Based on each product’s public docs as of November 2026. See the comparison pages for details.

Questions

Pricing is per workspace, not per seat: free for one repo, $19/mo for unlimited. See plans.

Do my prompts or code get uploaded?+

No. The log is parsed by JavaScript in this tab. If you press Save, only the numbers go to the server: token counts, timestamps, model ids, tool names and a short hash of each edited file name (so loops can still be detected). Never prompt text, responses, code or file names.

I am on a Claude Max or Pro subscription. Is this useful?+

You will see what the session would cost at API list prices, which is the right number for comparing tasks and models, but not what you pay. AgentReceipt matters most when you pay per token: the API, Bedrock, Vertex, or claude-code-action in CI.

Where are my session files?+

Claude Code: ~/.claude/projects/<project>/<session-id>.jsonl (the newest file is the latest session). Codex CLI: ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl. In GitHub Actions, claude-code-action writes an execution file whose path is in its execution_file output.

How accurate is the price?+

Each model call is priced from the token counts the agent itself logged, at the lab’s published list price for input, output, cache writes (5-minute and 1-hour) and cache reads. Negotiated discounts and Batch pricing are not applied.

How are the savings estimated?+

Each flag has a rule: a cache miss could have been a cache read, a loop’s extra turns could mostly have been avoided, a trimmed tool output would not be re-read. Savings are allocated per step and never exceed what a step cost, so the avoidable total never double counts.