Browser control

September 6, 2026 · View on GitHub

The dashboard has a Browser panel, and the agent can drive the page shown in it. Ask it to open a site, click through a flow, fill a form, or take a screenshot of what it found, and you watch the whole thing happen in that panel.

You can take over at any point with your real mouse and keyboard. That is how a CAPTCHA or a two-factor prompt gets handled: the agent stops, you do the step, it carries on.

What the agent can do to a page

Navigate to a URL, read the page as a structured list of its elements, click, type, press a key, hover, choose from a dropdown, take a screenshot, wait for something to appear, go back, and read the browser console.

Reading a page it can already reach is cheaper than driving it, so a request that only needs the text of a public page may be answered without the browser at all. Driving it is for the cases that need interaction, a logged-in session, or a page whose content only exists once its scripts have run.

Local addresses are refused

The agent cannot navigate the panel to localhost, a loopback address, or a private-network address. Those are where your own control planes live — this dashboard among them — so driving one would let the agent reach an interface it is not supposed to operate. Only globally routable addresses are accepted.

Settings → Browser

Two things live there:

  • Install the browser engine. Driving a page needs a browser binary. The panel installs it for you and shows whether it is present.
  • Attach to your own browser. Attach mode drives your everyday browser, with your live logins and your open tabs, instead of a fresh one. It needs a browser extension that only you can install — the panel links it — and an optional token stored here removes the per-attach approval prompt.

Treat an attached browser as borrowed. The agent should not navigate a tab away from what you were doing, and closing it would take your own windows with it.

  • Dashboard: the side panel and where its tabs live
  • Computer use: driving native desktop apps rather than a web page
  • Artifacts: keeping a screenshot or a page the agent produced