The built-in browser

A multi-tab browser in the right panel that the assistant can open, read, and screenshot — and, in Build and Fix, navigate, click, type, hover, scroll, and fill forms with a visible cursor.

Point it at any site, or at your own dev-server URL to run and test your app as you iterate. The agent drives it, or you do.

The default read returns the page content AND its interactive elements together, each with a stable reference the agent can act on — one call, so it never has to read twice to act.

To judge how a page LOOKS, it uses a text report of the rendered layout: blank or broken areas, invisible and low-contrast text, overlap, and layout shift. That is cheaper and more accurate for layout facts than an image.

A screenshot is a last resort, for when a literal picture is genuinely needed — a design comp to reproduce, an image's actual content, a canvas.

The agent can list, create, switch, and close tabs. It creates a new tab rather than navigating away whenever the current page still matters — a signed-in session, a filled form, search results to come back to.

Every tab shares one browser session, so if a site needs a login you do not have, signing in yourself carries over to the agent's tabs.

Browser console errors, uncaught exceptions with stack traces, and failed (4xx/5xx) requests surface in the Errors tab and are readable by the agent.

That is how it fixes what actually broke at runtime rather than guessing from the source.

In Build and Fix, the agent can run a non-interactive asserted test against your running app — navigate, act, assert — and get a single pass/fail plus any console errors and failed requests.

A turn will not be reported as finished while your app is throwing runtime errors.