Browser Skill

Provides server-side headless browser automation using agent-browser CLI (Playwright/Chromium backend):

• Navigate to URLs (with auto-snapshot and bot detection)
• Snapshot accessibility tree with element references (e.g., @e5)
• Click elements by reference
• Type text into input fields
• Scroll up/down
• Navigate back in history
• Press keyboard keys
• Take full-page screenshots (base64 PNG)
• Execute JavaScript in page context
• Read console messages and JS errors
• Session isolation per task_id

Use builtin_browser tool with fields:

• operation: one of "navigate", "snapshot", "click", "type", "scroll", "back", "press", "screenshot", "eval", "console", "close"
• url: URL to navigate to (required for navigate)
• ref: Element reference from snapshot (e.g., @e5) for click/type
• text: Text to type (for type) or key to press (for press)
• js: JavaScript expression to evaluate (required for eval)
• task_id: Session identifier (default: "default"); each task_id gets its own browser session
• direction: Scroll direction "up" or "down" (default: "down")
• full: Full page screenshot "true"/"false" (default: "false")
• clear: Clear console/errors after reading "true"/"false"

Response format:

• Always return the result in a single message. When taking a screenshot, embed the base64 PNG inline in your reply via ![alt](data:image/png;base64,...) — do not split into multiple messages.

Notes:

• Requires agent-browser CLI installed: npx agent-browser install --with-deps
• Each task_id maintains an isolated browser session with persistent state
• Sessions auto-clean after 5 minutes of inactivity
• Element references from snapshot (e.g., @e5) are used for click/type — not CSS selectors
• For simple HTTP fetches, prefer builtin_http_client (faster, lighter)