OpenAI’s Codex Computer Use gives AI agents desktop control capabilities, but it’s closed-source — you can only call the API, cannot understand internal mechanisms, and cannot train your own model.
trycua/cua is changing this. The project has rapidly climbed GitHub Trending this week, gaining 1,842 stars in a single week to reach 15,405 total stars.
Project Positioning: Complete Computer-Use Agent Infrastructure
cua is not another agent framework — it is the underlying infrastructure for Computer-Use:
Three Core Components
1. Cross-Platform Sandbox
| Platform | Support | Description |
|---|---|---|
| macOS | Yes | Full mouse/keyboard/screen control |
| Linux | Yes | X11 and Wayland support |
| Windows | Yes | Native desktop control |
2. SDK
- Unified screen observation interface: screenshots, DOM tree, accessibility tree
- Standard operation interface: mouse movement, click, keyboard input
- State management: Agent can track current operation state and context
- Multi-Agent support: Multiple agents can operate the same desktop simultaneously
3. Evaluation Benchmarks
- Standard task set: covering file operations, web browsing, application control
- Success rate metrics: quantifying agent performance across tasks
- Efficiency metrics: token consumption, operation steps, error rate
Comparison with Competitors
| Dimension | trycua/cua | OpenAI Codex CU | Claude Computer Use |
|---|---|---|---|
| Open source | Yes | Closed API | Closed API |
| Cross-platform | macOS/Linux/Windows | macOS/Linux | macOS |
| Multi-Agent | Yes | No | No |
| Custom training | Yes | No | No |
| Benchmarks | Yes | No | No |
Action Recommendations
- Agent framework developers: Evaluate cua SDK and sandbox for adding Computer-Use capability
- Researchers: Use cua benchmarks to compare different Computer-Use models
- Enterprise security teams: cua sandbox design provides a template for safe Computer-Use Agent deployment
- Individual developers: cua offers more flexibility and lower cost than closed-source API alternatives
In the Computer-Use domain, open-source infrastructure is catching up with closed-source APIs. trycua/cua represents this trend.