Overview
Cua is an open-source platform designed to help researchers and engineers build, benchmark, and deploy AI agents that can use and control full desktop environments. It focuses on safe, reproducible, and self-hostable infrastructure: isolated sandboxes for execution, SDKs for agent and computer control, and benchmark environments for evaluation and RL training.
Core Components
- Cua: The main agent SDK and runtime for creating computer-use agents. It exposes APIs to control virtual desktops, interact with UI elements, and run isolated code execution environments.
- Cua-Bench: A benchmark and RL environment suite (OSWorld, ScreenSpot, Windows Arena, etc.) for evaluating agent performance and exporting trajectories for model training.
- Lume: A macOS/Linux virtualization manager for Apple Silicon that uses Apple's Virtualization.Framework to run near‑native VMs.
Key Features
- Sandboxes & Isolation: Support for Docker, QEMU, and Apple Vz to run agents and untrusted code in isolated, self-hostable environments.
- Multi-OS Support: Target macOS, Linux, and Windows desktops so agents can learn to operate real-world applications and GUIs.
- SDKs & CLIs: Agent SDKs (cua-agent), computer control SDKs (cua-computer), server drivers (cua-computer-server), CLI tooling and image/VM management via Lume.
- Benchmarks & Training: Cua-Bench provides standardized tasks and environments suitable for reinforcement learning, evaluation, and dataset/trajectory export for offline training.
- Self-hostable & Privacy-friendly: Designed to be run on private infrastructure rather than relying on hosted cloud services.
Usage Examples
Simple Python example to create a computer and run an agent (requires supported models and SDK):
from computer import Computer
from agent import ComputerAgent
computer = Computer(os_type="linux", provider_type="cloud")
agent = ComputerAgent(model="anthropic/claude-sonnet-4-5-20250929", computer=computer)
async for result in agent.run([{"role": "user", "content": "Open Firefox and search for Cua"}]):
print(result)Intended Audience & Use Cases
- AI engineers building agents that interact with GUIs or need to run code in isolated environments.
- Researchers benchmarking agentic behavior and training models with exported trajectories.
- System engineers and developers who want to run macOS/Linux VMs on Apple Silicon using an integrated toolchain.
Licensing & Ecosystem
Cua is distributed under the MIT License. Some optional components may have their own licenses (e.g., AGPL for certain optional dependencies). The project provides documentation, examples, a blog, and a community Discord for contributions and support.
Why it matters
By providing reproducible sandboxes, standardized benchmarks, and agent SDKs, Cua lowers the engineering overhead for building agents that can operate real desktop software—bridging a gap between RL/LLM research and practical, interactive agent applications.
