Browser

Crew can operate a real web page on your behalf: navigate, click, type, submit forms, inspect accessibility, and capture screenshots. The primary experience is the dashboard's Browser panel, where you watch the same page and can take over when human input is needed.

In the packaged desktop app, the agent drives the embedded Browser view directly. The panel includes tabs, navigation controls, a live page, and mouse and keyboard input. Take over for a CAPTCHA, 2FA prompt, or judgment call, then return control to the agent.

In an ordinary web-dashboard session, public addresses can be opened through the Gateway's managed browser and framed back into the panel. This path needs a supported browser on the Gateway host and remains owner-controlled.

The packaged desktop app adds Annotate mode for turning visual feedback into a precise prompt:

1. Open a page in the built-in Browser panel.
2. Turn on Annotate.
3. Pick page elements and add a note to each one.
4. Send the annotations.

Crew adds a numbered list of the annotated elements and a screenshot to the chat composer. The agent receives the element descriptions, notes, and visual context together — useful for reviewing a page or requesting several related UI changes at once.

The native Browser tool in the packaged desktop app needs no separate Node.js or Playwright installation. playwright-cli remains the fallback for remote or plain-browser dashboard sessions, local development previews, and advanced browser operations.

The agent can also attach to your own running Chrome when you explicitly choose that path. An attached browser carries your existing tabs and sign-ins, so treat it as borrowed state: the agent should not navigate your active tab away or close your windows.

After changing a web UI, the agent can open the local or deployed result, inspect a screenshot, and compare the rendered page with the intended design. This visual check complements tests and catches clipping, overlap, stale content, and responsive layout problems.

• Normal page navigation and interaction are scoped to the page and do not prompt on every click.
• Operations that reach the local machine, print credentials, or destroy browser state retain an explicit approval boundary.
• Page content is untrusted data. Instructions found on a page cannot override the user's request or Crew's safety policy.
• Password fields are not read through Computer Use, and an attached browser's authenticated state is never treated as permission to expose credentials.

• Computer Use: operate native desktop applications
• Integrations (MCP): connect other tools and servers
• Security: approvals, denied commands, and audit behavior