agent-browser

What This Skill Does

The agent-browser skill is a high-performance, headless browser automation tool specifically engineered for AI agents. Unlike standard Puppeteer or Playwright wrappers, agent-browser focuses on accessibility-first interaction. By leveraging accessibility tree snapshots, it allows agents to interact with web elements using unique, deterministic 'refs' (e.g., @e1, @e2). This significantly reduces the fragility of automated scripts that rely on brittle CSS selectors or XPaths, which often change during site updates. It provides a full suite of navigation, interaction, state checking, and network control capabilities, making it ideal for complex workflows that require authentication, state persistence, and SPA navigation.

Installation

To install this skill, run the following command in your terminal:

clawhub install openclaw/skills/skills/matrixy/agent-browser-clawdbot

Ensure you have the OpenClaw environment initialized. You can verify the installation by running agent-browser --help after installation completes.

Use Cases

• Automated Web Testing: Run regression suites on complex single-page applications (SPAs) without dealing with UI element shifting.
• Data Scraping & Extraction: Extract large volumes of data from websites that require user interaction, such as clicking 'load more' buttons or handling infinite scrolls.
• Multi-session Workflows: Use session isolation to manage different user roles (e.g., 'admin' and 'user') simultaneously in separate browser instances.
• Task Automation: Automate repetitive manual web tasks like filling out forms, handling authentication cookies, or navigating through multi-step procurement or signup processes.

Example Prompts

1. "Open https://github.com/login, fill in the credentials using the accessibility tree to find the input fields, and save the session state to auth.json once logged in."
2. "Navigate to the dashboard and wait until the network is idle, then capture a snapshot of the main content and get the text from all list items with the class .data-row."
3. "Block all requests matching '**/ads/*' and then proceed to scrape the prices from the product page by clicking the next button until we reach the end of the list."

Tips & Limitations

• Accessibility Matters: Because this skill relies on the accessibility tree, ensure the target websites have proper ARIA labels for the best experience. If an element is missing from the snapshot, it might be hidden from accessibility tools.
• Performance: For large pages, use the -c (compact) flag during snapshots to keep the context window manageable for your LLM.
• Persistence: Always use state save and state load to avoid redundant login flows, which saves time and prevents triggers of anti-bot systems.
• Headless Mode: This skill runs headless by default. While this is efficient, some sites with sophisticated bot detection may require additional headers or user-agent spoofing, which can be configured via network routing tools.