Skip to main content
The embedded browser is a tool panel, the same rank as the terminal and the document preview. It is backed by the system Chrome, Chromium, or Edge installation and runs in the runtime’s own isolated profile. It is not an application shell and not another WebUI product. You and the Agent see the same page: you watch a live frame, while the Agent works from a structured page snapshot and a set of browser tools. That is what keeps the element it clicks and the element you see from diverging.

​
Open the browser panel

Choose Browser from the preview panel’s + menu, or create a new browser tab. The address bar supports:
  • A full address, opened as written.
  • A keyword, searched with the engine selected in Settings → Browser.
  • Back, forward, reload, and opening links in a new tab.
Browser tabs share the same tab strip and splits as files, diffs, and terminals. The panel can be popped out into its own window or shown full screen.
The embedded browser needs Chrome, Chromium, or Edge on the machine. When no browser is available, the panel says what is missing instead of showing a blank page.

​
Read and operate a page

Page content is untrusted input. The Agent’s tool descriptions tell it not to follow instructions embedded in a page; apply the same judgment when you approve navigation, uploads, or sensitive actions.

​
What the Agent can do

The Agent uses built-in browser tools:
  • Observe: read the accessibility tree and interactive elements, returning short-lived element references.
  • Act: click, fill, press keys, hover, scroll, select options, drag, and upload files.
  • Navigate and manage tabs: open, switch, close, and wait for a condition.
  • Read: extract page text, read console and network diagnostics, and take screenshots.
  • Test: record actions and export a Playwright test; save a page baseline and compare changes.
  • Ask for help: hand a captcha, sign-in, or undecidable moment back to you.
Element references are valid only until the next observation. After a rerender, navigation, or another observation, old references are refused and the Agent must observe again. This prevents “the click landed on whatever moved into that position”.

​
You can always take over

When the Agent is driving the current tab, the panel shows Agent is driving in the bottom-right corner. Press pause to take control immediately. While paused the Agent can still observe, but cannot click or type. Press play to hand the tab back; the Agent observes the page again before acting. Your own input into the page also pauses the Agent on that tab. The panel explains why and offers a hand-back action.
Browser actions can submit forms, send messages, incur charges, or change external systems. Confirm the site, account, and scope before approving. Never give an Agent passwords, payment details, or one-time codes.

​
Settings

Settings → Browser controls the browser panel’s own tabs:
With a remote runtime, the channel that carries the panel’s frames and display settings is not complete yet, so display options apply to the next local runtime. Local capabilities such as file choosers do not cross runtimes.

​
Security boundary

  • The browser uses its own user-data-dir; it does not read or write your everyday browser profile unless you explicitly import login state.
  • Downloads are denied by default and can only land in the runtime downloads directory, never in the workspace.
  • A page cannot silently open a native save dialog or file chooser.
  • Every navigation and action is written to a redacted audit ledger; cookies, tokens, and typed values are not.
  • Cross-origin navigation, loopback addresses, and uploads may require extra approval.

​
Troubleshooting