Skip to content

you probably got here from a file someone sent you

That file is a receipt.

It is the record of two coding agents checking each other's work. One of them wrote the change. The other one — a different model, from a different company — read the diff afterwards and went looking for bugs in it. The receipt is what it found, and the exact input that would have broken.

Nothing on that page came from us. It was written on the sender's own machine, by a tool they ran themselves, and it made no network request when you opened it — no pixel, no beacon, no identifier in the link you clicked. You are here because you chose to click, which is the only way we ever find out a receipt was read.

Why two agents and not one

A model reviewing its own work grades its own homework. Two models from two vendors do not share a blind spot, and the one reviewing has no stake in defending the code. Exactly one of them can write at a time — enforced by real permission flags, not by asking nicely — and then they swap.

What a finding has to clear

Style opinions are dropped unread. A finding only counts if the reviewer can name inputs or state that produce a wrong result, a crash, or data loss. That is why the receipt you were sent shows a concrete failing case next to each one rather than an adjective.

What it can't do

Trying it yourself

Not yet, and there is no command here that would pretend otherwise. ccx is not published and the repository is private, so nothing you could copy off this page would run. When it ships it will need Node 18.17 or newer plus the claude and codex CLIs signed in to your own subscriptions. It drives them on your own machine under your own logins, and it never resells model access or sees your code.

The docs are the full account of what it does. Receipts covers what is in the file you were sent, and what is scrubbed out of it first.

What this is →