cua

Cua is open-source infrastructure for Computer-Use Agents (CUA), with the goal of giving AI agents a usable computer and enabling them to operate a full desktop like a human. The project introduces the concept of "Computer-Use 2.0," describing agents switching between code, APIs, and graphical interfaces within the same task. Official materials position Cua as the party providing computers and automation tools; users can bring their own agents and models, or use its dedicated CUA-S1 decision model.

1. Main components and uses
2. Cua Spaces: A desktop application that provides agents with a full desktop. Each desktop is called a Space and can be a macOS virtual machine built on a local Mac, or a Linux or Omarchy image. Spaces can run on a Mac, other machines you own, and your own cloud accounts (AWS, Google Cloud, or Modal), and the application keeps them in the menu bar and notch area. Spaces support Teleport, which can move logged-in applications (such as Chrome or Slack) into a Space and keep them logged in there; sessions are stored in Cua Keyvault, encrypted on the Mac, and only enter a Space after user approval. It also supports Multiplayer, allowing users and agents to each have their own cursor on the same desktop and work collaboratively; users can step in to make choices and then hand the desktop back. Spaces images come with cua-spacesd built in, so once a Space starts, agents immediately gain process, file, screenshot, and input capabilities.
3. Cua Driver: Provides agents with tools to inspect and operate native desktop applications and browsers on macOS, Windows, and Linux, and can be connected via CLI, MCP, or typed SDKs. Where applications and platforms support it, background delivery allows agents to work without moving the user's pointer or stealing focus. The official tutorial example is having an agent calculate 6 × 7 in a calculator and verify that the application displays 42. There are three access methods—CLI, MCP, and SDK—and it can be integrated with agents such as Claude Code, Codex, Cursor, and OpenClaw.
4. Lume: Uses Apple's Virtualization.Framework to create and manage native [text truncated]macOS and Linux virtual machines. The official tutorial example is to create a native macOS Tahoe virtual machine from an Apple recovery image, boot it, and connect via SSH.
5. Cua SDK and CLI: Provides a unified SDK and cua command for local virtual machines and containers, as well as any machine running cua-spacesd. The SDK offers the same API in Python (pip install cua), TypeScript (@trycua/cua), Swift (Cua), and Kotlin (generated bindings), and can be embedded to run in-process or run through a shared cua daemon. The installation script displays a short checklist: the cua CLI, the Cua Spaces app (default on macOS), the cua-driver MCP and skill for agents, and the managed local machine; then run cua auth login to log in, and optionally install cua skills and the cua MCP server into AI coding agents (Claude Code, Codex, Cursor, etc.). You can preselect with sh -s -- --select cua-driver, or skip the checklist with --only cua-driver. The sandbox itself does not need a built-in agent; readiness comes from the runtime and optional port probing; images that ship with cua-spacesd (port 3211) gain process, file, screenshot, input, and low-latency audio/video streaming capabilities through cua-driver. cua host setup can make the local machine reachable through the cua.ai relay without port forwarding.
6. CUA-S1: A family of small, specialized System 1 models for computer use. "System 1" is used here as an engineering analogy, referring to fast, bounded decisions, such as choosing the value that should be filled into a field, or whether to leave an element unchanged; it is not a strict model architecture classification, nor does it replace the planning and reasoning of a general-purpose agent. The first research profile focuses on forms: scoring decisions from structured interface elements and document values, rather than generating responses token by token. The application code is responsible for ordering actions, and the optional Cua Driver integration handles execution within explicit action boundaries. The project includes Python model code, synthetic data generation, training, and evaluation; the GitHub portion is an early source-only research release, with model weights hosted separately on Hugging Face, the source code under the MIT license, and each model and dataset card should be checked separately for its scope, limitations, and specific license.
7. Cua Bench: Used to build computer-use tasks, evaluate agents, and export trajectories for training. The official recommendation is to start with simulated tasks that do not require a VM, Docker, or a model API key, requiring Python 3.12 or 3.13 and uv; the tutorial goal is to create a small task, run the reference solution, and verify that the evaluator reports a reward of 1.0.

II. Installation and Usage Cua Spaces must be installed on macOS 26 or later. The installer selects the Cua Spaces app by default and adds the cua CLI; you can also download the signed .dmg, or get the .pkg from the Cua Spaces 0.1.0 release. Cua Driver provides installation methods such as Windows PowerShell, and the official tutorials cover platform setup, permissions, and agent connection. The Lume tutorial uses the Lume CLI directly and explains the unattended installation defaults. SDK installation examples include pip install cua.

III. Licensing and Restrictions Most content outside Cua Spaces and Cua Driver Bench is MIT licensed, including the cua SDK and its Python, TypeScript, Swift, and Kotlin packages, the cua command and cua daemon, Cua Driver, Lume, and more. Cua Spaces uses the FSL-1.1-MIT source-available license, covering the Spaces app, cua-spacesd, Cua Keyvault, teleport, Cua Volume, as well as streaming clients, codecs, and viewers; the streaming wire protocol remains MIT. This license allows free use, self-hosting, and building on top of it, but you may not provide a competing hosted service, and each version converts to MIT two years after release. Cua Driver Bench is also FSL-1.1-MIT, with versions released before October 11, 2026 (#4973, #4978) under MIT. The MIT parts never depend on the FSL parts. Some subdirectories have their own licenses, and LICENSING.md lists each directory and the relationship between the two; putting SpWhether aces or Cua Driver Bench is offered as a hosted or managed service requires consulting COMMERCIAL.md; use of names and logos is covered in TRADEMARKS.md. Third-party components include: cua-som is an optional package licensed under AGPL-3.0-or-later, and its Ultralytics dependency retains its own license; the contents of the Microsoft OmniParser repository are CC-BY-4.0, and model files downloaded from a separate model repository are different artifacts; the optional cua-perception extension is installed separately from signed assets in the cua-perception-v<version> GitHub release, combining AGPL-3.0-only OmniParser model artifacts, Apache-2.0 PP-OCR model artifacts, and a separately packaged ONNX Runtime. This extension is not MIT-licensed, and redistributing it or providing it to users over a network may trigger AGPL-3.0 source code obligations; Cua does not relicense this detector. Cua Spaces is free for individuals, with Pro and Teams plans coming soon. Apple, macOS, Ubuntu, Canonical, and Microsoft are trademarks of their respective owners, and this project is not affiliated with or endorsed by these companies.

IV. Applicable Scenarios Cua is suitable for training, evaluating, and generating computer-use agent data, building agent desktop environments across operating systems, desktop automation and form decision-making research, and tasks that require agents to operate native applications and browsers. Before use, note the platform support boundaries, permission requirements, and the different license terms of each component.