Browser-Use Skill

Build AI agents that autonomously browse the web using LLMs and the Chrome DevTools Protocol.

When to Use

• Building AI agents that need to interact with web pages
• Automating web tasks (form filling, data extraction, navigation)
• Creating autonomous web research agents
• Testing web applications with AI-driven exploration
• Scraping dynamic content that requires JavaScript execution

Core Concepts

Architecture

User Task → LLM (Claude/GPT) → Browser-Use Controller → Chrome CDP → Web Page
                ↑                        ↓
                └──── Page State (HTML/Screenshot) ────┘

Installation

pip install browser-use
playwright install chromium

Basic Agent

from browser_use import Agent
from langchain_anthropic import ChatAnthropic

agent = Agent(
    task="Find the latest Python release version on python.org",
    llm=ChatAnthropic(model="claude-sonnet-4-20250514"),
)

result = await agent.run()
print(result)

Advanced Configuration

from browser_use import Agent, Browser, BrowserConfig

browser = Browser(
    config=BrowserConfig(
        headless=True,
        disable_security=False,
        extra_chromium_args=["--no-sandbox"],
    )
)

agent = Agent(
    task="Navigate to example.com and extract all links",
    llm=llm,
    browser=browser,
    max_actions_per_step=5,
    use_vision=True,  # Use screenshots for better understanding
)

Key Patterns

1. Multi-Step Task Automation

agent = Agent(
    task="""
    1. Go to github.com
    2. Search for 'browser-use'
    3. Click on the first repository result
    4. Extract the star count and description
    """,
    llm=llm,
)

2. Data Extraction

agent = Agent(
    task="Go to news.ycombinator.com and extract the top 10 story titles with their URLs",
    llm=llm,
    max_actions_per_step=3,
)

3. Form Interaction

agent = Agent(
    task="""
    Go to the contact form at example.com/contact
    Fill in: Name='Test User', Email='test@example.com', Message='Hello'
    Submit the form
    """,
    llm=llm,
)

Best Practices

1. Be specific in task descriptions — Clear, step-by-step instructions yield better results
2. Use vision mode for complex UIs — use_vision=True sends screenshots to the LLM
3. Set reasonable action limits — Prevent infinite loops with max_actions_per_step
4. Handle authentication carefully — Never hardcode credentials in task descriptions
5. Use headless mode in production — headless=True for server deployments
6. Implement error handling — Wrap agent runs in try/except for graceful failures

Common Pitfalls

• CAPTCHAs: Browser-Use cannot solve CAPTCHAs; use authenticated sessions instead
• Rate limiting: Add delays between actions for sensitive websites
• Dynamic content: Use use_vision=True for SPAs with heavy JavaScript
• Memory usage: Close browser instances after use to prevent memory leaks

References

• Browser-Use GitHub
• Browser-Use Documentation
• Chrome DevTools Protocol