Execute a browser task described in natural language. The AI agent navigates and interacts with pages automatically. Resilient to site layout changes — no CSS selectors needed.

Args: task: What to do, e.g. "Search for 'AI news' and return top 3 results" url: Optional starting URL to navigate to first max_steps: Maximum interaction steps (capped at 50) session_id: Optional session ID to reuse an existing browser session

Extract structured data from a webpage using natural language.

Args: url: Target URL data_description: What to extract, e.g. "all product names and prices in JSON" max_steps: Maximum interaction steps (capped at 50) session_id: Optional session ID to reuse an existing browser session

Close a persistent browser session.

Args: session_id: The session ID returned when the session was created

List active persistent browser session IDs.

Close all active persistent browser sessions.

Save a session's cookies + localStorage (auth state) to a JSON file.

The file holds live authentication secrets — store it securely and never commit it. Restore it by setting BROWSER_STORAGE_STATE= before the session starts.

Args: session_id: An existing session ID (create one via browse/extract with session_id) path: File path to write the storage-state JSON to

Manage browser tabs within a persistent session.

Args: session_id: An existing session ID action: One of "list", "open", "close" url: URL to open (action="open"; validated http/https) target_id: Tab target_id to close (action="close"; from action="list")

Open a URL and return a snapshot of the page's interactive elements.

Keyless — the calling AI drives the browser directly, with no server-side LLM. Follow up with snapshot/click/type_text/scroll/press_key to interact, and get_content to read the page as markdown. Stealth (Patchright) is applied.

Args: url: URL to open (http/https only) session_id: Persistent session to use (default: "default")

Return the current page's interactive elements as [index] <tag> text.

Pass an index to click/type_text. Re-run after the page changes.

Args: session_id: Session to inspect (default: "default")

Click the element with the given index (from snapshot). Returns a new snapshot.

Args: index: Element index from snapshot session_id: Session to act on (default: "default")

Type text into the element with the given index. Returns a new snapshot.

Args: index: Element index from snapshot text: Text to type session_id: Session to act on (default: "default")

Scroll the page, then return a new snapshot.

Args: direction: up, down, left, or right amount: pixels to scroll (default 500) session_id: Session to act on (default: "default")

Press a key or chord (e.g. "Enter", "Control+a"), then return a new snapshot.

Args: keys: Key or chord to send session_id: Session to act on (default: "default")

Return the current page as clean markdown (no LLM, no screenshot).

Args: session_id: Session to read (default: "default")

List files downloaded in this session, with their paths.

Files are saved to ./downloads by default (override with BROWSER_DOWNLOADS_PATH). Call this after a click/navigate that triggers a download to find the saved file.

Args: session_id: Session to inspect (default: "default")

Take a screenshot of a webpage and return it as base64-encoded PNG.

Keyless — navigates and captures directly, no LLM.

Args: url: URL to screenshot full_page: Capture the full scrollable page (default: visible viewport only)

Render a webpage to PDF and return it as a base64-encoded PDF.

Keyless — navigates and renders directly, no LLM. Headless only (Chromium prints PDF only in headless mode).

Args: url: URL to render