Browser session: status (mode, headless, browser version, profile, tabs, active tab), launch (restart with new options; tabs are not carried over, cookies persist in a persistent profile), close (close the browser in launch mode, disconnect in attach modes).

Tabs: list ("t1 "Title" url (active)"), new (optionally at url, becomes active unless background:true; returns a viewport snapshot when url is given (call snapshot for the whole page)), select (make active; brings it to the front in headed mode), close (default the active tab; the previously active tab becomes active). Tab ids are never reused; popups are registered automatically.

Go to a URL (scheme-less gets https://), or back/forward/reload. Returns "[tN] navigated "" ()" and a viewport snapshot (call snapshot for the whole page; snapshot:false to skip, snapshot:true attaches the whole page). HTTP error statuses are reported, network failures are errors with the net::ERR code. A page that loaded (committed) but did not reach wait_until within timeout_ms is not an error: the result adds "(still loading: …)"; only a navigation that never committed is an error, and it is stopped.

Compact accessibility snapshot of a tab: one line per element "- role "name" eN attrs", indented by nesting; iframes are inlined. interactive mode (default) lists controls, headings, landmarks, dialogs, alerts; full adds text, lists, tables, images. Page with offset (line offset) when the footer says so.

Screenshot of the viewport (default), the full page, or a target element. Default jpeg quality 70, long edge at most max_edge (1280). Afterwards, "x,y" targets use this image's pixel coordinates. annotate draws each visible ref label on its element (removed before returning). save writes it under /screenshots/.

Main content of the page (or of target) as markdown: title, headings, paragraphs, lists, links as text, simple tables. Navigation, footers, cookie banners, sidebars and ads are excluded. Long content is cut at max_chars content chars and ends with "… more chars (offset= to continue)"; pass that offset for the next page.

Find elements by case-insensitive substring over accessible name, visible text (also text split across elements: the smallest element holding it), placeholder, label and value, in all frames. Exact matches first. An empty query with a role lists every element of that role. Each hit is a snapshot line with its ref, followed by a line " in ".

Click a target (ref, "x,y", or selector). Auto-waits for the element to be visible, stable, enabled and not covered. Returns an effect summary (navigation, new tabs, dialogs, downloads, live-region announcements as → alert (not JS dialogs), console errors).

Type text into a field (input, textarea, contenteditable). Replaces the value by default; append adds to it; slowly sends key-by-key events (autocomplete widgets); submit presses Enter afterwards. The typed text is never echoed.

Fill several fields in one call and optionally click submit (only if every field succeeded). Values: text-like fields take a string (replaced); select takes an option value or label (string[] for multiple); checkbox/switch take a boolean; radio takes true, or a label/value on any radio of the group; file inputs take path(s) inside the allowed roots; date/time/color/range take the string value. One result line per field; values are never echoed.

Press keys or chords in order, e.g. ["Tab"], ["Control+a"], ["Meta+Shift+K"], ["ArrowDown","Enter"]. target is focused first.

Move the mouse over a target (reveals :hover menus and tooltips).

Scroll the page or a container. Without direction/to, target is scrolled into view. With direction (amount px, default 80% of the viewport/container) or to top|bottom, target is the scroll container (default the page). Reports "scrolled y=/ viewport=" and a height change (infinite scroll).

Drag from one target to another with real mouse events; works for HTML5 drag-and-drop and pointer-event drags.

Upload files: target is an input[type=file] (files set directly) or an element that opens a file chooser (intercepted; no native dialog appears). Paths must resolve inside the allowed roots (cwd, output dir, ~/Downloads, ~/Desktop, ~/Documents, --allow-file-root); dot-files are refused unless --allow-hidden-files.

Accept or dismiss the pending JS dialog (confirm, prompt, beforeunload) on a tab; text answers a prompt. Alerts are auto-accepted.

Wait for exactly one condition: text (appears), text_gone, target with state (visible default, hidden, attached, detached), url (substring or /regex/), load_state, or ms (unconditional, at most 30000). On timeout the error names the condition.

Run JavaScript in the page. script is an expression or a function source; a function is called with the target element (or no argument). Promises are awaited. The result is JSON (truncated); DOM elements are rendered as snapshot lines with refs. Page exceptions are returned as errors.

Watch videos on the page (all frames, incl. cross-origin iframes). list: one line per video (ref, frame, src, duration, time, paused, muted, volume, rate, size, readyState, text tracks). play/pause/seek(time)/set(rate, muted, volume, loop). frames: images by element screenshot, each preceded by "frame i/n @ mm:ss.ss": times=[s…] (seek each), count (default 6, evenly spaced over start..end), or every_ms + duration_ms (sample while playing); seek-based capture restores time and paused state; all-black frames (DRM) are flagged. captions: text-track cues as "mm:ss.ss-mm:ss.ss text", optional start/end window. Default target: the largest visible playing video, else the largest visible one.

Console messages and uncaught page errors of a tab, buffered from tab creation (at most 500): "[level] hh:mm:ss.mmm message (url:line)". level error|warning (warnings and errors)|all; pattern is a substring or /regex/; clear empties the buffer after reading.

Requests of a tab (at most 1000): "r METHOD status|failed: type url ms ". Filter by pattern (substring or /regex/), type (document, fetch, xhr, script, …), failed_only (network failures and HTTP >= 400). request_id shows headers (credentials redacted) and the body (text truncated at 10k; binary shows its size).

Run up to 50 tool calls sequentially in one call: steps [{tool, args}]. Each step's output is under "## step : " (images kept). stop_on_error (default true) lists the remaining steps as skipped after a failure. Over the output cap, every step keeps its header and first line and the rest is shared; a cut step ends with "… chars cut from step ". batch cannot nest.