Cua

Cua: sandboxed computers for AI agents, from one VM to a fleet

Cua (as in Computer-Use Agents) provides the infrastructure layer for agents that operate real computers. A single API boots sandboxed machines across Linux, Windows, macOS, and Android, spanning six local runtimes and cloud providers — so the same agent code that clicks through a Windows app locally can run against a fleet of cloud machines for training, evaluation, or production automation.

The platform is built for the workloads that make computer-use agents hard at scale. Snapshot-native rollouts fork machine state with copy-on-write storage, letting you run parallel episodes from an identical starting point or reproduce a failure exactly. Warm machine pools keep pre-booted machines claimable in milliseconds and scale to zero when idle. The open-source Cua Driver gives agents the primitives — click, type, scroll, inspect accessibility trees — while MCP server and CLI surfaces plug the whole thing into existing agent stacks.

Key Features

• Cross-OS fleet management: one API for Linux, Windows, macOS, and Android machines across six local runtimes and cloud providers
• Snapshot-native rollouts with copy-on-write forking for parallel episodes and exact failure reproduction
• Warm machine pools claimable in milliseconds, scaling to zero when idle
• Open-source Cua Driver (MIT licensed) for click, type, scroll, and accessibility-tree inspection
• Verified trajectory data: human-reviewed golden trajectories with step-level annotations for agent training
• Deployment flexibility: free open-source stack, hosted cloud, BYOC, or on-premises, with SOC 2-ready posture

Ideal Use Case

Cua fits two audiences: research and model teams that need thousands of reproducible OS environments for training and evaluating computer-use models, and product teams shipping agents that must drive real desktop software — legacy Windows apps, native macOS tools, Android flows — beyond what browser-only automation reaches. Start free on the open-source stack, then move to hosted or BYOC fleets when concurrency and compliance requirements grow.

How Cua differentiates

Cua is Y Combinator-backed, MIT licensed, and its GitHub repository has over 19,000 stars; its site notes Google DeepMind evaluating Gemini computer-use capabilities on the platform, alongside users like Qwen Code. Browser-sandbox providers such as Browserbase stop at the web, and code-sandbox platforms like E2B stop at the terminal — Cua's scope is the entire operating system across four OS families, with snapshot forking and warm pools designed specifically for RL-style parallel rollouts.

FAQ

Is Cua open source? Yes — the core stack and Cua Driver are MIT licensed on GitHub, with a free tier to start.

Which operating systems are supported? Linux, Windows, macOS, and Android, bootable through a single API across local runtimes and cloud providers.

What are snapshots used for? Forking machine state with copy-on-write storage — run many parallel agent episodes from one starting point, or replay a failure exactly.

How do I deploy at scale? Options include Cua's hosted cloud, bring-your-own-cloud, and on-premises deployments; dedicated fleets are available by request.

tl;dr

Cua provides open-source, sandboxed virtual machines across Linux, Windows, macOS, and Android — with snapshots, warm pools, and an SDK — for training, evaluating, and running computer-use agents.

Related

Looking for more options? Browse the Developer Tools directory or read our best AI coding tools listicle. Cua is also tracked on Crunchbase.