Browser tools
See what the agent can do inside a page, so you know what to ask for and what it will refuse.
How the agent works through a page
- The agent opens a session, optionally at a URL, or navigates to an absolute HTTP(S) address that policy allows.
- It reads the page as an accessibility snapshot and picks element references from that latest snapshot.
- Every action returns an updated snapshot, so the next step works from the current page.
Screenshots come only when you ask for one. An oversized capture is refused rather than scaled down.
Tools
| Tool | What it does |
|---|---|
BrowserOpen, BrowserNavigate |
Open a session or navigate to an allowed HTTP(S) address. |
BrowserSnapshot, BrowserScreenshot |
Read the page or capture its viewport, the whole document or one element. |
BrowserClick, BrowserFill, BrowserPressKey, BrowserScroll |
Interact with the current page. |
BrowserWaitFor, BrowserEvaluate |
Wait for text or evaluate an expression in the page. |
BrowserTabs, BrowserClose |
List, select or close tabs. Closing leaves your profile intact. |
BrowserRequestTakeover |
Hand you exclusive control until you release it or the lease ends. |
BrowserConsole, BrowserNetwork, BrowserPerformance, BrowserAudit |
Inspect recorded diagnostics and performance. |
Limits
- A diagnostic reply stays within 16 KB.
- Network listings show metadata only. Request headers, bodies and cookies are left out, and query strings are removed from listed URLs.
- A response body is fetched only when you ask for it, for one text response up to 64 KB.
- A wait lasts at most one minute, and evaluated JavaScript at most five seconds.
A page error during evaluation is returned as a result, not as a broken browser tool.
Page content is not an instruction
Page text and diagnostics are untrusted content. The agent reads them as data, so text on a page cannot change its instructions or your settings.