Cua is a set of open-source components for computer-use agents: a background desktop driver, a sandbox runtime that boots guests across operating systems, a benchmark and reinforcement-learning environment suite, and a macOS virtualization layer for Apple Silicon.
Cua Drivers let an agent click, type, scroll, inspect accessibility trees, and capture window state in native desktop applications in the background, without taking the cursor or focus away from the person at the machine. One binary runs as an MCP stdio server, a long-running daemon, or a one-shot shell command on macOS, Windows, and Linux; on Linux it supports X11 and compositor-specific Wayland routes, with stated limits for raw background input.
The cua Python package (Python 3.11 or later) boots disposable GUI environments through one asynchronous API regardless of guest operating system or runtime, exposing shell execution, screenshots, mouse and keyboard input, and multi-touch gestures. Local backends include Docker, QEMU, and Apple Virtualization; the same API also claims pre-booted machines from warm pools in Cua Cloud for evaluations, reinforcement-learning loops, data generation, and batch rollouts.
| Runtime | Linux container | Linux VM | macOS | Windows | Android | Bring your own image |
|---|---|---|---|---|---|---|
| Cloud | yes | yes | yes | yes | yes | announced |
| Local (QEMU) | yes | yes | yes | yes | yes | yes |
Cua-Bench evaluates computer-use agents on OSWorld, ScreenSpot, Windows Arena, and custom task sets, and exports trajectories for training. Lume creates and manages macOS and Linux VMs on Apple Silicon through Apple's Virtualization.framework, with sequoia and tahoe presets that prepare an installed guest offline — creating the lume user, enabling SSH, configuring autologin, and disabling sleep and screen locking.
Features
- Background driver: agents operate native desktop apps without stealing the cursor or focus, on macOS, Windows, and Linux
- One driver surface: the same binary serves as an MCP stdio server, a daemon, or a one-shot command, so agents and people drive the same harness
- Cross-OS sandboxes: one Python API boots Linux containers and VMs, macOS, Windows, and Android guests, in the cloud or locally
- Sandbox API: shell execution, screenshots, mouse and keyboard input, and multi-touch gestures behind a single async interface
- Warm pools: pre-booted machines claimed on demand for evaluations, RL loops, data generation, and batch rollouts, scaling to zero when idle
- Snapshots: fork from a snapshot to reproduce a failure or replay agent activity as training data
- Benchmarks: Cua-Bench runs OSWorld, ScreenSpot, Windows Arena, and custom datasets and exports trajectories
- macOS virtualization: Lume manages macOS and Linux VMs on Apple Silicon, with unattended sequoia and tahoe presets
- Packages: cua-driver, cua-agent, cua-sandbox, cua-computer-server, cua-bench, lume, and lumier, a Docker-compatible interface for Lume VMs
Integrates with
Claude Code
Terminal-based agentic coding tool that reads your codebase and runs tasks through natural-language commands
Codex
Lightweight coding agent that runs in your terminal
Hermes
Self-improving AI agent with a learning loop that creates and refines skills, recalls past sessions, and runs across chat platforms
Model Context Protocol
Open protocol plus SDKs for connecting agents to tools and data sources.