Subagent Browser Skill

This skill documents the usage of the Subagent Browser. It is the primary, default tool for navigating the web, interacting with complex UIs, and bypassing automated bot detection.

[!TIP] The subagent browser natively maintains a persistent Chrome user profile. This means cookies, active login sessions (like Google accounts), and site preferences persist across multiple invocations!

I. Architecture & Setup

• Tool Name: default_api:browser_subagent (or simply browser_subagent).
• Browser Instance: It runs a headed, visible Chromium process.
• Video Recording: The subagent automatically records its entire session. When it finishes, a WebP video file is permanently saved to your current conversation's artifacts directory. You do not need to write custom Python scripts to record the screen.
• Bot Detection: This browser perfectly bypasses advanced headless bot detection scripts (like those used by Cloudflare or sannysoft), making it ideal for interacting with highly secure websites.

---

II. Usage Guidelines

Because the subagent operates as an autonomous miniature AI entity, you do not feed it individual tool calls (like click or type). Instead, you invoke it with a comprehensive task prompt.

1. Be specific: Tell it exactly what to look for and what data to return in its final report.
2. Define completion: Explicitly state the condition that should cause the subagent to stop and return (e.g., "Wait for the table to load, then read the first 5 rows and return").
3. Use RecordingName: Provide a descriptive 2-3 word name for the WebP video recording it generates.

---

III. Internal Tool List (The Escape Hatch)

The subagent comes equipped with 17 highly specialized browser interaction tools.

[!NOTE] DYNAMIC DISCOVERY: The list below was inventoried in June 2026. If you find yourself needing a capability that isn't listed, or if the subagent errors out during execution, you can instruct the subagent to run a self-diagnostic (e.g., Task: List all your available tools and their schemas) to dynamically discover newly added or updated commands.

Current Capabilities Inventory

• list_browser_pages: Lists all open pages in the browser and their metadata.
• capture_browser_console_logs: Retrieves the console logs of an open browser page.
• browser_move_mouse: Moves the mouse cursor to a specific pixel coordinate on the browser page.
• browser_press_key: Simulates keyboard key presses or text input on the active browser page.
• browser_resize_window: Resizes the browser window or changes its state (e.g., maximized, fullscreen).
• browser_scroll: Scrolls the browser page at a specific coordinate by horizontal and vertical deltas.
• browser_mouse_down: Presses and holds a mouse button at the current cursor position.
• browser_mouse_up: Releases a pressed mouse button at the current cursor position.
• capture_browser_screenshot: Captures and optionally saves a screenshot of the viewport or a specific element.
• click_browser_pixel: Performs a click (left, right, or double) at a specific pixel coordinate. (Highly useful for React/Canvas UIs that block standard DOM clicks).
• browser_drag_pixel_to_pixel: Simulates dragging the mouse from a start coordinate through waypoints to an end coordinate.
• read_browser_page: Reads the visible text, title, URL, and window dimensions of an open browser page.
• browser_get_dom: Retrieves the interactive elements and text from the DOM within the current viewport.
• execute_browser_javascript: Runs custom JavaScript code within the context of an open browser page.
• open_browser_url: Navigates the browser to a specified URL, optionally replacing an existing page.
• browser_list_network_requests: Lists network requests made by a browser page.
• browser_get_network_request: Retrieves detailed information about a specific network request by its ID.