Browser

Crew can drive a real web browser on your behalf: open a page, click, type, fill forms, and capture the page. This reaches work that only exists behind a web UI — checking a deployed site, walking a web console, or reviewing the agent's own front-end changes.

Browsing is powered by Playwright. The agent runs the playwright-cli browser CLI as ordinary shell commands, and Crew owns the install and the dashboard surface that shows the live session.

The dashboard's Browser panel embeds the browser's own live view: a session grid with a live screencast, a detail view with a tab bar and navigation controls, and full mouse and keyboard input so you can take over a session directly. That handover is the path for a CAPTCHA or a 2FA prompt the agent can't and shouldn't complete — drive it yourself, then hand it back. Press Escape to release input capture.

The agent can navigate, click elements, type into fields, submit forms, and snapshot the page. Ask in chat — "open the staging site and check the login page" — and the panel follows along.

Browsing becomes available once the playwright-cli browser CLI is installed; there's no separate on/off switch. Crew installs it for you:

• It checks for playwright-cli and Node.js 20 or newer on your PATH.
• If it's missing, Crew installs the CLI and downloads the browser binary, giving you a progress surface rather than a stall on first use.

Because installing the browser is what grants the capability, having it installed is treated as your consent to let the agent browse. The Settings → Browser surface documents this and is where browser settings live.

Sessions are named, so an agent can keep a logged-in context separate from a throwaway one. The agent can also attach to your own running Chrome, operating a browser that already holds your logged-in sessions — the most powerful form of the capability, and worth being deliberate about, since the agent then acts with your existing sign-ins.

When the agent changes a web UI, it can open the result in the browser and judge what it sees, so a front-end change is verified visually in the same turn rather than described and hoped for.

Browser actions run as shell commands, so they're governed by the same permission model as any other command:

• Page-scoped actions run without a prompt — navigating, clicking, typing, and reading the page don't interrupt you.
• Actions that reach your machine keep interactive approval — for example writing a screenshot to a path, or evaluating arbitrary script.
• Page content is treated as untrusted. A URL, instruction, or form target read off a page never decides the agent's next navigation, so a malicious page can't steer the session.

The live-takeover view is bound to loopback only; it's never exposed to the network, because it carries full remote input over a browser that may hold your sessions.

• Computer Use — drive native desktop apps, not just the browser
• Integrations (MCP) — connect other tools and servers
• Security — the permission model, denied commands, and audit log