CUA (github.com/trycua/cua) is an open-source Computer-Use Agent infrastructure project providing sandboxes, SDKs, and benchmark tooling for training and evaluating AI agents that can control full desktops (macOS, Linux, Windows). The project has accumulated 15,100 stars, 943 forks, 3,199 commits, and 761 branches.
The Problem It Solves
The core challenge for Computer-Use Agents isn’t “making AI click on screens” but training and evaluating these agents in controlled, safe, reproducible environments. CUA provides this infrastructure layer:
- Sandbox Environments: Isolated desktop environments where agents can safely execute operations
- SDK: Unified programming interface for building and integrating CUA agents
- Benchmarks: Standardized evaluation frameworks for comparing agent performance on desktop tasks
- Cross-platform Support: Covers macOS, Linux, and Windows
The project recently added GitHub Copilot CLI MCP integration, indicating alignment with mainstream developer tooling.
Project Architecture
cua/
├── cua-driver/ # Desktop driver layer
├── cua-sdk/ # Developer SDK
├── cua-sandbox/ # Sandbox environments
├── cua-bench/ # Benchmark suite
├── examples/ # Example projects
└── docs/ # DocumentationQuick Start
git clone https://github.com/trycua/cua.git
cd cua
pip install -e cua-sdk/
docker compose up -d cua-sandbox
python examples/basic_desktop_agent.py