cua

Cua is an open-source project that aims to give AI agents computers they can operate. It bundles several components under one repository, covering desktop automation, sandboxed environments, virtual machines, specialist models, and evaluation tooling. The project describes its scope as "Computer-Use 2.0," meaning an agent that can move between code, APIs, and graphical interfaces within a single task. Main components: - Cua Driver: tools for inspecting and operating native desktop apps and browsers on macOS, Windows, and Linux. It can be reached through a CLI, MCP, or typed SDKs. Background delivery is offered so agents can act without moving the pointer or stealing focus, where the app and platform allow it. Install scripts are provided for macOS/Linux (shell) and Windows (PowerShell). - Cua Fleets: isolated cloud desktops provisioned through run.cua.ai. A fleet keeps sandbox capacity, and code claims a desktop from a pool and uses the Sandbox SDK to run commands, capture screenshots, and interact with apps. Local sandboxes and fleets share the SDK but differ in credentials, images, operations, and runtime requirements. - Lume: creation and management of local macOS and Linux VMs on Apple Silicon using Apple's Virtualization.Framework, with a CLI and an install script. - CUA-S1: a family of small, specialized "System 1" models for bounded computer-use decisions, such as choosing a value for a field. The first research profile focuses on forms, scoring decisions from structured interface elements and document values rather than generating tokens. The repository contains Python model code, synthetic-data generation, training, and evaluation; it is described as an early, source-only research release with weights hosted separately on Hugging Face. - Cua Bench: tooling to build computer-use tasks, evaluate agents, and export trajectories for training. A simulated task can be run without a VM, Docker, or a model API key, using Python 3.12/3.13 and uv. The README points to documentation, tutorials, a blog, Discord, and GitHub Issues. It states the project is MIT-licensed, with third-party components under their own licenses (for example Kasm, an optional AGPL-licensed cua-som package, and OmniParser-related artifacts). Trademark notices clarify no affiliation with Apple, Canonical, or Microsoft. Citation metadata is provided via BibTeX and CITATION.cff.