Cua

Building a Manus-style agent that clicks, types and drives apps on its own is hard not because of the model, but because it needs a safe computer and tools that can read and control desktop programs. Cua open-sources that foundation: a computer-use framework that lets AI agents operate computers like people do, so you don’t reinvent the wheel.

A driver layer reads and operates native apps and browsers on macOS, Windows and Linux; sandboxes provide isolated cloud desktops or local VMs; and specialized small models plus benchmarks round it out. Bring your own model while Cua supplies the computer and the tools.

Features

Cua Driver:Inspect and operate native apps and browsers on three platforms via CLI, MCP or SDKs, in the background without stealing your pointer.

Sandboxes:Local sandboxes and cloud desktop fleets share one Sandbox SDK for commands, screenshots and app control.

Lume:Spin up macOS and Linux VMs on Apple Silicon for clean agent environments.

CUA-S1:Small models specialized for computer-use decisions, paired with your own LLM.

Cua Bench:Create tasks, evaluate agents and export trajectories for training.

Code plus GUI:Agents move between code, APIs and interfaces within one task.