Cua

Cua (from Cua AI, Inc., a Y Combinator company in San Francisco) gives AI agents computers they can actually use. Where most agent tooling stops at the browser, Cua covers the whole desktop — native applications on macOS, Windows, Linux and Android — and it is built as five separable pieces rather than one monolithic agent, so you bring your own model and orchestration and Cua supplies the machine and the controls.

Cua Driver is the part an agent talks to. It exposes click, type, scroll and accessibility-tree inspection through an MCP server, a background daemon or a CLI, with typed SDKs alongside. Its distinguishing trick is background delivery: where the platform and application allow it, the agent operates the app without stealing your pointer or window focus, which is the difference between an automation you can watch and one that locks you out of your own laptop. Integration guides exist for Claude Code, Codex, Cursor and OpenClaw. It installs with a single shell or PowerShell command.

Cua Fleets is the hosted side: pools of isolated cloud desktops that keep warm capacity so code can claim a machine instead of waiting for one to boot, driven through the same Sandbox SDK as a local sandbox. Fleet pricing is metered at $0.044625 per vCPU-hour and $0.0223125 per GB-hour; running sandboxes locally through Docker, QEMU or Apple's Virtualization.framework costs nothing.

Lume creates and manages local macOS and Linux VMs on Apple Silicon using Apple's Virtualization.framework — the practical answer to "I need a real, disposable macOS box" for agent work. Cua Bench builds computer-use tasks, evaluates agents against them and exports trajectories for training; a simulated task runs with no VM, Docker or model API key at all.

CUA-S1, announced in September 2026, is Cua's family of small "System 1" models for computer use: fast, bounded decisions such as which value belongs in a field or whether to leave an element alone, scored from structured interface elements rather than generated token by token. The first research profile targets forms. The source is an early MIT-licensed research release with weights and the training dataset published separately on Hugging Face; the repository also ships the synthetic-data generation, training and evaluation code.

The open-source core is MIT-licensed on an open-core model, with on-premise deployment and SOC 2 Type II available for teams that need them. It is a good fit for anyone building or evaluating computer-use agents at scale — RL environments, data generation, cross-platform QA — and overkill for someone who only needs a headless browser.