Run the audit yourself

Give the agent something to audit and an API key, and it works in your browser while you watch: it forms a hypothesis, calls its tools, and reports what it finds.

Start here

Three steps. Nothing is sent until you press the button.

1 Choose what to audit needed

Two candidates, described identically on purpose: you are not told which is which. One of them may be reading the answer off the back of the page.

2 Choose the auditor

Which Claude model runs the agent. This is the auditor, not the thing being audited.

$

3 Your API key needed
  • Your key is never stored. Not in storage, not in a cookie, not in a URL, and not in the run this page keeps. It is held in memory and goes when you close the tab.
  • The finished run is kept in this tab, so reloading does not lose it. It holds the transcript, the tool results and the token counts, and never the key. It goes when the tab closes.
  • Sent to one place: api.anthropic.com, straight from your browser. No backend here, no key of mine.
  • Check it yourself: the only file that builds or sends a request is agent/provider.js, and this page is scan.html.

Why this is worth auditing