CUA addresses a critical infrastructure gap in the AI agent ecosystem: how to safely give AI agents the ability to interact with operating systems and desktop applications without compromising host security or sacrificing performance. The platform combines Apple's Virtualization.Framework with agent-friendly interfaces to create lightweight VMs that run at 97% of native CPU speed on Apple Silicon. Each sandbox is ephemeral and reproducible, ensuring agent tests run in identical conditions every time. The unified Computer SDK abstracts away OS-specific details, letting developers write agent logic once and deploy it across macOS, Linux, Windows, and Android environments through Docker, QEMU, or native virtualization.
The platform consists of three core components that form a complete development lifecycle. CuaBot is a multi-agent CLI that lets developers run any AI agent inside a sandbox with H.265 streaming and shared clipboard support. The Cua Agent SDK connects frontier AI models from Anthropic, OpenAI, Google, Microsoft, and Alibaba — or local models via Ollama and LM Studio — to virtual computers through observe-reason-act loops with budget limits and trajectory recording. Cua-Bench provides standardized benchmarks from OSWorld, ScreenSpot, and Windows Arena, plus tools for creating custom evaluation tasks and exporting trajectories for reinforcement learning training. Developers can run thousands of agent trajectories in parallel across hundreds of sandboxes simultaneously.
CUA is developed by a Y Combinator X25 company and has been trusted by over 50,000 engineers. The platform offers both self-hosted and cloud deployment options, with cloud sandboxes providing hot-start under one second through snapshot and restore functionality. MCP server integration allows CUA agents to be used as tools in Claude Desktop, Cursor, or any MCP client. The project is MIT licensed with usage-based cloud pricing starting at a free tier, with Pro plans from $10 per month billed per resource: CPU at $0.0145 per core per hour, memory at $0.0033 per GiB per hour, and disk at $0.000044 per GiB per hour.