Cua (by trycua) is an open-source infrastructure suite designed for Computer-Use Agents—AI models that can control a full desktop environment just like a human would. It provides the "body" (drivers and sandboxes) for an AI "brain" (like Claude or GPT-4) to see the screen, move the mouse, and type across macOS, Windows, Linux, and Android.
A unified SDK (pip install cua) to manage ephemeral environments. It allows developers to spin up isolated virtual machines (VMs) or containers where agents can safely execute tasks without affecting the host machine.
A background driver for macOS and Windows that allows agents to drive native apps without "stealing" the user's actual cursor or focus.
Includes Lume, a specialized tool for running macOS VMs on Apple Silicon with near-native performance, which is essential for testing Apple-specific workflows.
A "co-op" tool that lets agents like Claude Code or Cursor run inside a sandbox while projecting their windows onto your desktop via H.265 streaming.
Cua-Bench provides a framework for running reinforcement learning (RL) environments and exporting "trajectories" (recordings of agent actions) for model training and evaluation.
- AI Researchers: Evaluating models on complex GUI-based tasks (OSWorld, Windows Arena).
- Software Engineers: Building "AI employees" that need to use legacy desktop software.
- QA Automation Teams: AI-driven cross-platform application testing.
pip install cuafrom cua import Cua
# Initialize a sandbox
with Cua().sandbox(os="macos") as sb:
# Tell the agent to do something
sb.display.screenshot().save("screen.png")
sb.keyboard.type("Hello World")
sb.mouse.click(x=100, y=200)Cua is an open-source project. Please refer to its repository for licensing details.