Computer Use

Computer Use Skill

Overview

This skill provides Playwright-powered browser automation capabilities via MCP, enabling programmatic interaction with web pages including clicking, typing, scrolling, and screenshot capture.

Context Savings: ~90% reduction

• MCP Mode: ~12,000 tokens always loaded (6 tools + cursor overlay)
• Skill Mode: ~300 tokens metadata + on-demand loading

Requirements

• Python 3.10+ installed
• Playwright (auto-installed by executor)
• Chromium browser (auto-installed via playwright install chromium)

Tools

The server provides 6 tools for browser automation:

Lifecycle Tools

Interaction Tools

State Capture Tools

Quick Reference

# List available tools
python executor.py --list

# Initialize browser (headless)
python executor.py --tool initialize_browser --args '{"url": "https://example.com", "width": 1440, "height": 900}'

# Initialize browser (visible)
CU_HEADFUL=1 python executor.py --tool initialize_browser --args '{"url": "https://google.com"}'

# Click at coordinates (0-1000 scale)
python executor.py --tool execute_action --args '{"action_name": "click_at", "args": {"x": 500, "y": 300}}'

# Type text at coordinates
python executor.py --tool execute_action --args '{"action_name": "type_text_at", "args": {"x": 500, "y": 300, "text": "Hello World", "press_enter": true}}'

# Scroll to 50% of page
python executor.py --tool execute_action --args '{"action_name": "scroll_to_percent", "args": {"y": 500}}'

# Click by CSS selector
python executor.py --tool click_selector --args '{"selector": "button.submit", "nth": 0}'

# Fill form field
python executor.py --tool fill_selector --args '{"selector": "input[name=search]", "text": "Claude AI", "press_enter": true}'

# Take screenshot
python executor.py --tool capture_state --args '{"action_name": "after_click"}'

# Execute JavaScript
python executor.py --tool execute_action --args '{"action_name": "execute_javascript", "args": {"code": "document.title"}}'

# Close browser
python executor.py --tool close_browser

Configuration

Environment Variables

Setup

1. First run (auto-installs dependencies):

   python .claude/skills/computer-use/executor.py --list
   

2. For visible browser automation:

   export CU_HEADFUL=1
   export CU_SLOW_MO=500
   

Coordinate System

The system uses a 0-1000 normalized scale for viewport-independent automation:

• x=0 = left edge, x=1000 = right edge
• y=0 = top edge, y=1000 = bottom edge

This allows the same coordinates to work across different screen sizes.

Example: To click the center of the viewport:

{ "action_name": "click_at", "args": { "x": 500, "y": 500 } }

Tool Details

initialize_browser

Launch Playwright Chromium browser with specified configuration.

Parameters:

• url (string, required): Initial URL to navigate to
• width (number, optional): Viewport width in pixels (default: 1440)
• height (number, optional): Viewport height in pixels (default: 900)
• headless (boolean, optional): Run headless (overrides CU_HEADFUL env var)

Returns:

{
  "ok": true,
  "url": "https://example.com",
  "width": 1440,
  "height": 900,
  "headless": true,
  "slow_mo_ms": 250
}

execute_action

Execute a single browser automation action.

Parameters:

• action_name (string, required): One of:
  • open_web_browser - Navigate to URL
  • click_at - Click at coordinates
  • type_text_at - Type text at coordinates
  • scroll_to_percent - Scroll to position
  • press_key - Press keyboard key
  • execute_javascript - Run JS code
• args (object, required): Action-specific arguments

Action Arguments:

capture_state

Take a screenshot of the current browser state.

Parameters:

• action_name (string, required): Label for the screenshot file
• result_ok (boolean, optional): Whether previous action succeeded
• error_msg (string, optional): Error message if action failed

Returns:

{
  "ok": true,
  "path": "/tmp/gemini_computer_use/1234567890_action.png",
  "mime_type": "image/png",
  "url": "https://example.com"
}

click_selector

Click an element by CSS selector.

Parameters:

• selector (string, required): CSS selector for the element
• nth (number, optional): Index if multiple matches (default: 0)

fill_selector

Fill a form field by CSS selector.

Parameters:

• selector (string, required): CSS selector for the input field
• text (string, required): Text to enter
• press_enter (boolean, optional): Press Enter after typing
• clear (boolean, optional): Clear field before typing (default: true)

close_browser

Close the browser and release all Playwright resources.

Parameters: None

Usage Examples

Web Search Automation

# Initialize browser
python executor.py --tool initialize_browser --args '{"url": "https://google.com"}'

# Click search box (approximate center)
python executor.py --tool execute_action --args '{"action_name": "click_at", "args": {"x": 500, "y": 400}}'

# Type search query
python executor.py --tool execute_action --args '{"action_name": "type_text_at", "args": {"x": 500, "y": 400, "text": "Claude AI", "press_enter": true}}'

# Capture results
python executor.py --tool capture_state --args '{"action_name": "search_results"}'

# Close browser
python executor.py --tool close_browser

Form Filling with Selectors

# Initialize
python executor.py --tool initialize_browser --args '{"url": "https://example.com/form"}'

# Fill fields by selector
python executor.py --tool fill_selector --args '{"selector": "#name", "text": "John Doe"}'
python executor.py --tool fill_selector --args '{"selector": "#email", "text": "john@example.com"}'

# Click submit button
python executor.py --tool click_selector --args '{"selector": "button[type=submit]"}'

# Capture result
python executor.py --tool capture_state --args '{"action_name": "form_submitted"}'

JavaScript Extraction

# Get all links on page
python executor.py --tool execute_action --args '{"action_name": "execute_javascript", "args": {"code": "Array.from(document.querySelectorAll(\"a\")).map(a => ({text: a.textContent, href: a.href}))"}}'

# Get page title
python executor.py --tool execute_action --args '{"action_name": "execute_javascript", "args": {"code": "document.title"}}'

Integration with Agents

This skill integrates with the following agents:

• qa: For automated browser testing and visual verification
• developer: For debugging web applications
• devops: For deployment verification and smoke tests

Troubleshooting

Security Notes

• Browser automation executes locally via Playwright
• No remote code execution by default
• Screenshots saved to /tmp/gemini_computer_use/ (or system temp)
• Clear screenshots after sensitive operations