browser-use Skill for Claude Code

A wrapper around the official browser-use library that enables Claude Code to perform browser automation tasks through two modes: Direct Mode (Claude controls browser directly) and Subagent Mode (autonomous agent execution).

Features

• Direct Mode: Claude directly controls browser via Actor API (no external LLM API key required!)
• Subagent Mode: Delegate complex browser tasks to autonomous subagents
• Session Persistence: Server mode keeps browser session alive across multiple calls
• Full Automation: Navigate, click, type, screenshot, scroll, and more
• Bot Detection Bypass: Uses browser-use's stealth capabilities

Triggers

AI Automation Requests:

• "browse to...", "open website..."
• "search for...", "find on web..."
• "fill out form...", "automate..."
• "web research...", "scrape..."
• "take screenshot of..."

Development Testing:

• localhost URLs
• QA automation
• E2E testing

Installation

1. Install browser-use

pip install browser-use
playwright install chromium

2. Copy skill to Claude Code

cp -r browser-use-skill ~/.claude/skills/browser-use

3. Start server

cd ~/.claude/skills/browser-use
python server.py start &

Quick Start

Direct Mode (Recommended - No API Key Required!)

Claude directly controls the browser step by step:

cd ~/.claude/skills/browser-use

# Start server
python server.py start &
sleep 2

# Navigate to page
python server.py call '{"tool": "navigate", "args": {"url": "https://google.com"}}'

# Get page state with screenshot
python server.py call '{"tool": "get_state", "args": {"include_screenshot": true}}'
# Returns: elements list + screenshot_path (read with Vision!)

# Click element by index (from get_state)
python server.py call '{"tool": "click", "args": {"index": 0}}'

# Type text
python server.py call '{"tool": "type", "args": {"index": 0, "text": "search query"}}'

# Press Enter
python server.py call '{"tool": "press_key", "args": {"key": "Enter"}}'

# Take screenshot
python server.py call '{"tool": "screenshot", "args": {"path": "result.png"}}'

Subagent Mode

Delegate browser tasks to Claude Code subagents:

Use Task tool with subagent_type: "general-purpose"

Prompt template:
"Browser automation task using browser-use skill (Direct Mode).
Goal: [your task]
Server already running on port 9223.
Workflow: get_state -> analyze screenshot -> click/type -> repeat until done"

Server Commands

python server.py start     # Start server (use & for background)
python server.py stop      # Stop server
python server.py status    # Check status
python server.py call '{"tool": "...", "args": {...}}'

Tool Reference

Page Tools

Element Tools

Mouse Tools

Tab Management

AI Agent Tools (Requires API Key)

Examples

Example 1: Google Search (Direct Mode)

cd ~/.claude/skills/browser-use
python server.py start &
sleep 2

# Open Google
python server.py call '{"tool": "navigate", "args": {"url": "https://google.com"}}'

# Get elements
python server.py call '{"tool": "get_state", "args": {"include_screenshot": true}}'

# Type search query (index 0 is usually search box)
python server.py call '{"tool": "type", "args": {"index": 0, "text": "Claude AI"}}'

# Press Enter
python server.py call '{"tool": "press_key", "args": {"key": "Enter"}}'

# Screenshot results
python server.py call '{"tool": "screenshot", "args": {"path": "google_results.png"}}'

Example 2: Form Filling

# Navigate to form
python server.py call '{"tool": "navigate", "args": {"url": "https://example.com/form"}}'

# Get form elements
python server.py call '{"tool": "get_state", "args": {"include_screenshot": true}}'

# Fill fields (use indices from get_state)
python server.py call '{"tool": "type", "args": {"index": 0, "text": "John Doe"}}'
python server.py call '{"tool": "type", "args": {"index": 1, "text": "[email protected]"}}'

# Submit
python server.py call '{"tool": "click", "args": {"index": 5}}'

Example 3: Coordinate Click (When Index Fails)

# Get screenshot first
python server.py call '{"tool": "get_state", "args": {"include_screenshot": true}}'

# Analyze screenshot with Vision, then click at coordinates
python server.py call '{"tool": "mouse_click", "args": {"x": 500, "y": 300}}'

Example 4: Keyboard Navigation

# Tab through elements
python server.py call '{"tool": "press_key", "args": {"key": "Tab"}}'
python server.py call '{"tool": "press_key", "args": {"key": "Tab"}}'
python server.py call '{"tool": "press_key", "args": {"key": "Enter"}}'

# Keyboard shortcuts
python server.py call '{"tool": "press_key", "args": {"key": "Control+a"}}'  # Select all
python server.py call '{"tool": "press_key", "args": {"key": "Control+c"}}'  # Copy

Troubleshooting

"Server not running" error

python server.py start &
sleep 2
python server.py status

Element click not working

1. Re-fetch elements: get_state with include_screenshot: true
2. Check correct index in elements list
3. Try keyboard navigation: press_key("Tab") then press_key("Enter")
4. Use coordinate click: mouse_click(x, y)

Elements showing as "unknown"

This is normal for complex pages. The element cache still works - use the index to click/type.

Browser not visible

Server runs headless by default in background. Browser window appears but may be behind other windows.

Session disconnected

Always use server.py (server mode). Each server.py call maintains the same session.

How It Works

Why Server Mode?

Without server mode, each Python call would:

1. Launch browser
2. Perform action
3. Close browser (lose state!)

With server mode:

1. Start server once (browser stays open)
2. All calls share same browser session
3. Stop server when done

Direct Mode vs Subagent Mode

Direct Mode:

• Claude reads screenshot (Vision)
• Claude decides next action
• Claude calls tool
• Repeat until done

Subagent Mode:

• Claude spawns subagent with task description
• Subagent autonomously navigates
• Subagent reports results
• Better for complex multi-step tasks

Requirements

• Python 3.8+
• browser-use library
• Playwright + Chromium
• (Optional) LLM API key for run_agent/run_code_agent tools

License

MIT

DPA - Decentralized Protection Alliance

Freedom without surveillance, protection for everyone.

dpa.network | [email protected] | Matrix | Telegram