agent-browser

The agent-browser skill provides robust web navigation and interaction capabilities for AI agents, specifically designed for the NanoClaw containerized ecosystem. It allows agents to perform complex tasks by mimicking human web usage, including navigating sites, filling dynamic forms, clicking interactive elements, and extracting structured data. This skill is ideal for researchers, engineers, and power users who need to automate web-based workflows without exposing their host environment to external web content, as the entire browsing session is sandboxed within a Linux container.

• Full web navigation: open, back, forward, reload, and PDF generation.

• Intelligent snapshotting: generates accessibility trees and identifies interactive elements via DOM refs (e.g., @e1).

• Interaction suite: supports clicking, double-clicking, typing, checking boxes, selecting dropdown values, hovering, and uploading files.

• Information retrieval: extracts text, innerHTML, values, attributes, and page metadata, or counts specific elements.

• Semantic search: find and interact with elements using human-readable attributes like roles, labels, placeholders, or visible text.

• Session management: supports cookie operations, local storage manipulation, and saving/loading authentication states for persistence.

• JavaScript execution: runs custom scripts directly in the browser context for advanced page manipulation.

• Wait utilities: intelligent waiting for specific elements, network idle states, text appearance, or URL patterns.

• Best practice: Always perform a snapshot before attempting interactions to acquire current, valid element references.

• Efficiency: Use the interactive snapshot (-i) to limit the DOM tree scope and reduce token overhead during processing.

• Persistence: Leverage state saving for authenticated sessions to avoid repetitive login procedures.

• Isolation: The skill operates through secure Bash calls to the container environment, ensuring no direct access to the host machine's sensitive resources.

• Monitoring: For long-running tasks, use the screenshot feature to periodically capture the visual state of the browser for agent verification.