browser-automation

Browser Automation

AI-powered browser automation via browser-use MCP server. Navigate web pages, fill forms, extract content, take screenshots, and verify deployments.

Setup

Add to ~/.mcp.json:

{
  "mcpServers": {
    "browser-use": {
      "command": "uvx",
      "args": ["browser-use", "--mcp"]
    }
  }
}

Restart Claude Code after adding.

Usage

Navigate & Extract

/browser-automation navigate https://docs.example.com
/browser-automation extract https://docs.example.com --format markdown

Form Interaction

/browser-automation fill https://app.example.com/login
  email: test@example.com
  password: [from env TEST_PASSWORD]
  submit: button[type=submit]

Deploy Verification

/browser-automation verify https://myapp.com
  - Check: homepage loads (< 3s)
  - Check: /api/health returns 200
  - Check: login page renders
  - Screenshot: homepage, login, dashboard

Screenshot

/browser-automation screenshot https://myapp.com --full-page
/browser-automation screenshot https://myapp.com --element "#hero-section"

Integration Points

| Trigger | How It Helps | |---------|-------------| | shipper deploys | Auto-verify live URL, take screenshots | | e2e-runner needs browser | Natural language browser tests | | oracle needs deep docs | Navigate multi-page documentation | | designer needs references | Capture UI patterns from live sites | | qa-engineer tests forms | Fill and submit forms, verify results | | growth analyzes competitors | Extract features, pricing, UX patterns |

MCP Tools Available

| Tool | Description | |------|-------------| | browser_navigate | Go to a URL | | browser_click | Click element by selector or text | | browser_type | Type into input field | | browser_screenshot | Capture page/element screenshot | | browser_extract | Extract page content as text/markdown | | browser_wait | Wait for element or condition | | browser_evaluate | Execute JavaScript in page context | | browser_scroll | Scroll page or element |

Rules

• Max 1 request/second to same domain
• Respect robots.txt
• Never store credentials in output
• Timeout: 30 seconds per page
• Retry once on failure, then report error
• Blur sensitive data in screenshots