Opens a new browser session and returns its sessionId. The session (cookies, storage, page state) stays alive across tool calls until closed or idle-timed-out, so you can interact with the same page over many turns.

Lists open browser sessions with their current URL, age, and idle time.

Closes a browser session and frees its resources.

Navigates the session to a URL and returns the new page state plus a fresh snapshot with element refs. Relative URLs resolve against the session baseUrl.

Clicks an element. Address the element by ref (from the latest browser_snapshot), or by css, or by role+name. Reports any navigation, console errors, or network activity the click triggered.

Types text into an input or textarea. Address the element by ref (from the latest browser_snapshot), or by css, or by role+name.

Presses a key, optionally focused on an element. Key names follow Playwright ("Enter", "Escape", "ArrowDown", "Control+a").

Hovers an element, e.g. to reveal a menu or tooltip. Address the element by ref (from the latest browser_snapshot), or by css, or by role+name.

Selects one or more options in a . Address the element by ref (from the latest browser_snapshot), or by css, or by role+name.

Scrolls an element into view, or scrolls the page by a pixel delta when no element is given.

Waits for text to appear, text to disappear, or a selector to become visible. Use after actions that trigger async updates.

Navigates back in history and returns the new page state with fresh refs.

Answers an alert/confirm/prompt. A dialog blocks the page until it is answered, so an unanswered one is dismissed automatically rather than stalling the action that opened it. Call this BEFORE the action that triggers a dialog to arm the answer — that is the only way to accept one or supply prompt() text. Called while a dialog is open, it answers that dialog immediately.

Returns the accessibility tree of the page as compact YAML, with a [ref=eN] handle on every element. This is the primary way to see the page — use these refs to click and type. Far cheaper than HTML or screenshots. Scope it with ref/css/role, cap it with depth, or page through it with offset when a page is large.

Finds elements by role+name, visible text, or CSS, and returns a compact line per match including its ref and visible/enabled state. Use this instead of a full snapshot when you already know what you are looking for.

Returns the rendered text of the page or of one element subtree — the readable content without markup. Use it to verify copy, read results, or check an error message.

Captures a JPEG of the page or an element. Use this only for genuinely visual questions (layout, styling, rendering); browser_snapshot and browser_read_text are faster and cheaper for finding and verifying content.

Returns buffered console messages and uncaught page errors (with stacks). Check this whenever the page misbehaves — it usually names the failure directly.

Lists network requests the session has made, with status, size, and duration. Each line carries an #id for browser_request_detail. Filter by URL substring or by failure status.

Inspects one request from browser_network: headers, timing breakdown, request body, or response body. Small text responses are cached, so they stay readable after navigation.

Runs JavaScript in the page and returns the JSON-serialized result. Accepts a bare expression ("document.title") or a function ("el => el.value"). When an element is targeted, it is bound to el. Use it to read state the accessibility tree does not expose.

DevTools-style inspection of one element: tag, attributes, box model, form state, and computed styles. Use it to diagnose layout and visibility problems.

Delegates a multi-step UI task to a fast built-in browser agent, which drives the session itself and reports back. Use it for goal-shaped work ("log in as demo@example.com and check the dashboard loads", "walk the checkout flow and report anything broken") rather than driving each click yourself.

Returns a structured report: a success flag, a summary, and findings — each with a severity, what is wrong, where, and the evidence observed. Findings come from two places: what the agent noticed, and what the harness itself recorded (console errors, failed requests, dialogs), so problems are reported even when the agent does not mention them or runs out of steps. The session is left on whatever page the agent ended on, so you can inspect it further with the browser_* tools.