receipts

Your agent says “all tests pass.”
Check the receipts.

Coding agents end every task with a confident summary. receipts runs what they wrote and checks every claim against what actually happened: backed, contradicted or never ran.

Google Geminiis the agent: writes the code
Meta Muse Sparkreads the whole log, finds every claim
Subquadratic Snaprules on every claim in ~0.4 s

Four steps. Watch each one happen.

Every box lights up while its model is working, with a live clock. Nothing is pre-recorded: a run calls three sponsors' APIs and executes real code.

01 · Agent

Does the work

Google Gemini gemini-3.8-flash

Gets a ticket. Writes the code, its own tests, and the “done” message it would send you.

waiting
02 · Run

Actually runs it

Sandbox python -m unittest

Executes the agent's tests. This output is the only evidence anyone gets.

waiting
03 · Read

Finds every claim

Meta muse-spark-1.3

Reads the message and the whole log, lists each claim and quotes the exact log line behind it.

waiting
04 · Rule

Rules on each one

Subquadratic subq-snap-1.1

One call, every claim: backed, contradicted or never ran, each with a probability.

waiting
What the agent told you
Receipts, claim by claim
What actually ran
The code the agent wrote

Don't take the agent's word for it.

Every run below really happened, including real pull requests by GitHub's Copilot coding agent. Nothing is planted. Click one to replay it.

Add it to your agent.

receipts is also an MCP tool. Your agent calls it before it says “done”, learns which claims it can't back up, and fixes them first. Works in Claude Code, Cursor, Codex, Windsurf or any MCP client. No model keys needed.

Two tools your agent gets

receipts_run_and_check

Runs your tests (or any command), captures the real output, and checks the agent's “done” message against it.

receipts_check

Checks a message against a log you already have: a CI log, a terminal session.

What your agent sees (a real response)
receipts: 0 backed, 1 contradicted, 2 never ran.
- CONTRADICTED (100%) All 12 tests pass | log: "FAILED (failures=2)"
- NEVER RAN (99%) Added the parser | log: nothing
- NEVER RAN (100%) lint is clean | log: nothing
Fix or remove contradicted claims, and either verify or
soften never-ran claims, before reporting to the user.
Install

        
  1. Download receipts_mcp.py: one file, Python standard library only.
  2. Register it with the command above.
  3. Ask your agent to finish a task. Before it reports, it calls receipts and fixes what it can't prove.

Docs: docs/MCP.md. Self-host: set RECEIPTS_URL to your own deployment.

→ / space next · ← back · N notes · Esc exit