Browser Use

Browser Use

Use the browser Tool to carry out the user's website task and verify the result. Required browser components are prepared automatically on first use. Use the browser Tool for browser control; its permissions apply throughout the task.

Working loop

1. Start with open and the task's URL. To work in an already connected, logged-in page, use tabs and select the matching id with switch_tab.
2. Read the returned snapshot. Use its exact element refs for interaction. After page changes, obtain fresh refs with snapshot or request observe: true on the action. Refs from an earlier Run are no longer valid.
3. Fill related text and selection fields together in one fill call. Use {target, text} for text inputs and {target, text, kind: "select"} for option values in the fields array. An empty text clears a text input. Use press for a key combination. Request one observation after the group when you need a new target ref or confirmation. You can submit a focused field with press without an intervening snapshot. Limit requested observations with selector and limit when only one page section matters.
4. Verify the requested outcome using visible confirmation or page content. Continue until the task is complete or a concrete blocker remains; opening a page alone does not complete a task that asks for an action.

Reading and visual work

Use read for page text and follow next_offset when more is needed. A targeted snapshot with selector keeps large pages manageable; full: true includes noninteractive content. Use screenshot when layout or visual content matters. The image is returned directly. Use scroll to reveal more content and wait for expected visible text.

Tabs, dialogs, and files

Use ids returned by tabs when switching or closing tabs. Use new_tab for separate work. Handle a page dialog with dialog; accept only when appropriate to the task.

For downloads, activate the website's download control and use downloads to obtain completed file paths. For uploads, pass absolute paths on the browser's computer to upload. In a connected browser, downloads remain on that computer.

Recovery and completion

If refs are stale, take a new snapshot and continue. If input failed or its effect is uncertain, inspect the page before repeating it. A partial fill reports how many fields completed; continue from the observed state. An observation error does not mean the preceding action failed.

After submitting or navigating, check the returned URL and content before another input. If the page is still changing, use wait with expected visible text or an exact destination url; omit both when no specific condition is known. Wait returns a fresh observation. A quiet page or a completed input command does not prove that a form was submitted or the task succeeded.

Use status to inspect connection mode and headed without starting a browser. A managed browser window appears on the vBot server computer, which may differ from the user's computer. If a visible managed browser is needed, enable Show managed browser in the Browser Use Extension settings and reconnect; closing a managed connection discards its temporary browser profile. Attached browsers use their existing host and profile.

When a website shows an error, inspect its text and the reported HTTP status when available. Report that evidence; do not invent a cause such as automation detection. Avoid repeatedly submitting the same request to an unchanged error page. Preserve the user's requested website and method; report the concrete blocker when that required route remains unavailable.

Use existing logins when available. If the site requires a user-only login or Chrome displays a connection-consent dialog, identify that specific blocker. Summarize the verified result and provide relevant downloaded files. Use close when the browser connection is no longer needed; it disconnects a user-owned browser and closes a browser started for this Session.