🖥️ Desktop-Use

AI-powered desktop automation SDK — like Browser-Use but for your entire desktop.

from desktop_use import Agent

agent = Agent(task="Open Calculator and calculate 2+2")
result = agent.run_sync()

✨ Features

• 🤖 AI-Powered - Uses Claude's vision to understand and interact with any desktop application
• 🖱️ Full Control - Mouse, keyboard, screenshots - everything you need
• 🍎 macOS First - Native performance with cliclick (Windows/Linux coming soon)
• 📐 Ultra-wide Ready - Automatic resolution scaling for any display size
• 🧠 Memory Integration - Optional Remembra integration for persistent context
• ⚡ Simple API - One-liner to automate complex tasks

🚀 Quick Start

Installation

# Install cliclick (macOS only)
brew install cliclick

# Install Desktop-Use
pip install desktop-use

Grant Permissions

Go to System Settings → Privacy & Security → Accessibility and enable your Terminal app.

Run Your First Task

# CLI
desktop-use "Open Spotlight and search for Notes"

# Python
python -c "from desktop_use import run; run('Open Calculator')"

📖 Usage

Python SDK

from desktop_use import Agent, AgentConfig

# Simple usage
agent = Agent(task="Open FaceTime")
result = agent.run_sync()

# With configuration
config = AgentConfig(
    model="claude-4-sonnet-20250514",
    max_steps=20,
    verbose=True
)
agent = Agent(task="Create a new folder called 'Test' on Desktop", config=config)
result = agent.run_sync()

if result.success:
    print(f"Done: {result.final_message}")
else:
    print(f"Failed: {result.error}")

Low-Level Control

from desktop_use import Desktop

desktop = Desktop()

# Direct control
desktop.click(100, 200)
desktop.type("Hello World")
desktop.press("return")
desktop.hotkey("cmd", "space")  # Open Spotlight

# Screenshot
img = desktop.screenshot()
img.save("screen.png")

CLI

# Basic usage
desktop-use "Open Safari and go to google.com"

# With options
desktop-use --max-steps 25 "Fill out the form on the current page"
desktop-use --model claude-4-opus-20250514 "Complex multi-step task"
desktop-use --quiet "Background task"

🖥️ Platform Support

🔧 Requirements

• Python 3.10+
• macOS 12+ (for now)
• cliclick (brew install cliclick)
• Anthropic API key

📐 How It Works

1. Capture - Takes a screenshot of your desktop
2. Scale - Resizes to fit Claude's vision constraints (max 1568px edge)
3. Analyze - Claude sees the screenshot and decides what to do
4. Execute - Mouse/keyboard commands are executed via cliclick
5. Repeat - Loop until task is complete or max steps reached

The coordinate scaling is automatic - Claude works in scaled coordinates, and Desktop-Use maps them back to your actual resolution.

🧠 Memory Integration (Optional)

Connect to Remembra for persistent memory:

config = AgentConfig(
    memory_url="http://localhost:8787",
    memory_project_id="my-automation"
)
agent = Agent(task="Do what I asked yesterday", config=config)

🛣️ Roadmap

• macOS support with cliclick
• Ultra-wide display support
• CLI interface
• Windows support (pyautogui)
• Linux support (pyautogui/xdotool)
• MCP server for Claude Desktop
• Async operations
• Action recording/playback
• Visual debugging mode

🤝 Contributing

Contributions welcome! See CONTRIBUTING.md.

📄 License

MIT License - see LICENSE.

Built with ❤️ by DolphyTech