Open-source infrastructure for Computer-Use Agents providing sandboxes, SDKs, and benchmarks to train and evaluate AI agents that can control full desktops across macOS, Linux, and Windows.
One-Minute Overview#
Cua is an open-source platform for building, benchmarking, and deploying agents that can use any computer. It provides isolated, self-hostable sandboxes (Docker, QEMU, Apple Vz) that enable AI agents to interact with desktop environments, click buttons, and complete tasks autonomously.
Core Value: Provides complete computer control capabilities for AI agents without worrying about security isolation and environment compatibility issues.
Quick Start#
Installation Difficulty: Medium - Requires Python 3.12 or 3.13 environment with Docker for sandbox deployment
# Install Cua agent
pip install cua-agent
# Setup sandbox environment
cua sandbox setup --provider docker --os-type linux
Is this suitable for my scenario?
- ✅ AI Research: When you need to train and evaluate computer-use agents
- ✅ Development Tools: When building AI coding assistants or automated testing tools
- ✅ Cross-platform Automation: When executing complex tasks across different operating systems
- ❌ Simple Scripts: For basic task automation, this might be overly complex
Core Capabilities#
1. Computer Control SDK - Cross-platform UI Automation#
- Provides unified API for controlling different desktop environments (Windows, macOS, Linux)
- Supports screen recognition, mouse clicks, keyboard input, and basic interactions
- Includes file operations and application launching capabilities Actual Value: Developers can build AI agents that truly understand and operate computer interfaces, not just make API calls
2. Isolated Sandbox Environments - Secure Execution#
- Supports multiple virtualization technologies (Docker, QEMU, Apple Vz)
- Provides computation environments isolated from the host system
- Ensures security and reproducibility of agent operations Actual Value: Safely run AI agents without risking impact on the main system while ensuring consistent testing environments
3. Cua-Bench Benchmarking - Performance Evaluation Platform#
- Offers standard test suites including OSWorld, ScreenSpot, Windows Arena
- Supports reinforcement learning training environments and trajectory export
- Allows customization of evaluation tasks and metrics Actual Value: Objectively evaluate and compare performance of different computer-use agents, advancing the field
4. Lume Virtualization - High-Performance macOS/Linux VMs#
- Achieves near-native performance for macOS/Linux VMs on Apple Silicon
- Uses Apple Virtualization.Framework technology
- Provides Docker-compatible interface Actual Value: Get near-native performance for CI/CD, testing, and agent workloads without additional hardware
Tech Stack & Integration#
Development Languages: Python, Swift, HTML, TypeScript, Shell, Jupyter Notebook Key Dependencies: Requires Docker environment, supports various virtualization technologies (QEMU, Apple Vz) Integration Method: SDK / API
Maintenance Status#
- Development Activity: Actively maintained with multiple commits per week
- Recent Updates: Recent new releases available
- Community Response: Active community support through Discord and GitHub Issues
Commercial & Licensing#
License: MIT
- ✅ Commercial Use: Allowed
- ✅ Modification: Allowed
- ⚠️ Restrictions: Must include original license and copyright notice
Documentation & Learning Resources#
- Documentation Quality: Comprehensive
- Official Documentation: https://cua.ai/docs
- Example Code: Rich collection of example code available