Welcome

What it is

Eight checks. Six of them measure: they run in a browser on your machine, cost nothing, and need no API key. Two of them read and reason, and those send page text and screenshots to the model whose key you configured.

Everything one of them finds becomes a finding: a title in plain language, the surface it was found on, the evidence the browser captured, and a fix prompt you can paste into Claude Code, Cursor or Codex. You give each finding a verdict. The verdicts are what make the tool’s confidence numbers mean anything.

What it is not

• Not a linter. It presses things. A button that looks correct in the DOM and does nothing when clicked is invisible to static analysis and is the first thing this catches.
• Not a monitoring tool. It does not watch production and page you. It runs when you run it, or when your CI does.
• Not a replacement for real users. Test drive simulates people, and simulated people are for exploration. They find the door that does not open; they cannot tell you whether anybody wanted to walk through it.

The shortest path

1. Install it Node 20 or newer, then one browser download.
2. Answer the setup questions so it knows what your product is. This is not optional; without a profile every check stops and says so rather than guessing.
3. Run Speed & errors. It is free, it needs no key, and it usually returns something on the first pass.
4. Read what came back and give the first few a verdict.

That is about ten minutes, and none of it spends anything.

Where it runs

By default, on your machine. The console binds to the loopback interface, your findings are files in a directory you own, and there is no account. Hosting it for a team is a mode you turn on deliberately. See the notes on hosted mode.

It also runs where you already work: as an MCP server, so Claude Code, Codex and Cursor can start a run and read the findings without you leaving the editor.