Cua
Open-source infrastructure for Computer-Use Agents. Sandboxes, SDKs, and benchmarks to train and evaluate AI agents that can control full desktops (macOS, Linux, Windows).
Last verified:
What is Cua?
Cua is an open-source computer-use agent platform that gives AI agents full cloud desktops on Linux, Windows, macOS, and Android. It enables developers to build AI agents that can see screens, reason about what they're looking at, and interact with them through mouse, keyboard, and shell commands. The platform provides isolated sandbox environments where agents can execute tasks safely without risking the host system.
Key features include cloud desktops with hot-start under 1 second via Snapshot API, the Sandbox SDK for programmatic sandbox control, Agent SDK for vision-language model integration with 100+ VLM options, Cua-Bench for evaluating computer-use agents on real desktop tasks, and Lume for local macOS sandboxes on Apple Silicon. The platform supports Claude Code, Codex, OpenClaw, and other computer-use agents through its MCP Server integration with Claude Desktop and Cursor.
Cua is designed for AI coding assistants teams, browser and desktop automation developers, cross-platform testing engineers, and researchers benchmarking computer-use agents. Over 50,000 engineers at companies like Google, Meta, Apple, and NVIDIA use Cua for AI automation workflows, VPN automation, form filling, contact export, and complex UI navigation tasks.
The platform offers both cloud-hosted sandboxes and self-hosted options via Docker or Lume, with Python and TypeScript SDKs, REST API access, and support for multiple model providers including Anthropic, OpenAI, Google, Microsoft, and Alibaba through LiteLLM integration.
Cua pricing
Pricing model: Freemium
Cua offers simple usage-based pricing paying per minute of sandbox time. Free tier includes free sandbox minutes and a Linux Small cloud sandbox with full API and SDK access for local validation and first cloud tests. Pro plan starts from $10 with usage-based billing for active developers and teams, offering pay-as-you-go per-resource pricing for Linux plus Windows cloud sandboxes with priority support and trace visibility. Linux Sandbox per-resource rates: $0.0145 per CPU core/hour, $0.0033 per GiB memory/hour, $0.000044 per GiB disk/hour. Small sandbox (2 cores, 4 GiB, 50 GiB) costs ~$0.0444/hour or ~$32.41/month. Medium (4 cores, 8 GiB, 100 GiB) costs ~$0.0888/hour or ~$64.82/month. Large (8 cores, 16 GiB, 200 GiB) costs ~$0.1776/hour or ~$129.65/month. XLarge (16 cores, 32 GiB, 500 GiB) costs ~$0.3596/hour or ~$262.51/month. Enterprise plans available with custom pricing. Local Docker usage is free with no account needed.
Cua pros
- Open-source framework with MIT license
- Supports Linux, Windows, macOS, and Android sandboxes
- Hot-start sandbox in under 1 second with Snapshot API
- 100+ vision-language model options through VLM Router
- No account needed for local Docker usage
- Free API key available for cloud sandboxes
- 50,000+ engineers already using at major tech companies
- Python 3.12/3.13 SDK and TypeScript SDK available
- Prebuilt images designed specifically for computer-use agents
- MCP Server integration with Claude Desktop and Cursor
- Cua-Bench for verifiable task evaluation with oracle solutions
- Lume provides near-native speed on Apple Silicon
- Pay-as-you-go billing per minute of sandbox time
- Full root access in cloud desktops
- Support for Claude Sonnet 4.5, Opus 4.5, Haiku 4.5, Gemini 3
- Composed models available with grounding + planning
- Live GUI session via VNC access
- Interactive shell access to sandboxes
Cua cons
- Python 3.14 not supported due to pydantic-core compatibility
- Windows 10 Pro/Enterprise required for Windows Sandbox local option
- macOS local sandboxes require Lume CLI installation
- Golden image preparation needed for Linux and Windows Docker images
- Android images have fewer customization options
- Cloud sandbox requires API key that cannot be recovered if lost
- Usage-based pricing may become expensive for heavy workloads
- Linux images require ~2GB Ubuntu ISO download
- Windows images require ~6GB ISO and 90-day trial setup
- Learning curve for sandbox and agent SDK configuration
Frequently asked questions about Cua
What is Cua and what does it do?
Cua is an open-source sandboxes and agent framework for computer-use AI that lets you spin up isolated desktop environments on Linux, macOS, Windows, and Android, then build agents that can see and control them. It provides one SDK that works across any OS, either cloud or local, enabling AI agents to interact with real desktops through vision-language models.
How do I get started with Cua?
You can run locally with Docker without an account, or get a free API key at cua.ai for cloud sandboxes and hosted models. Install the Cua CLI, login with your API key, and create a sandbox. You can also use the website dashboard at Dashboard > Sandboxes > Create Sandbox to choose Linux, Windows, or macOS. From zero to a running agent takes under 5 minutes.
What operating systems does Cua support?
Cua supports Linux (Ubuntu), Windows 11, macOS, and Android 11 sandboxes. You can run these in the cloud via Cua Cloud, locally in Docker containers on any platform, on macOS VMs with Lume, or on Windows with Windows Sandbox. The same SDK works across all operating systems.
What model providers does Cua support?
Cua supports Anthropic (Claude models), OpenAI (GPT models including o1), Google (Gemini 3 Pro and Flash), Microsoft, and Alibaba models through LiteLLM. You get 100+ VLM options through Cua VLM Router, including composed models like GTA1 grounding with Claude Sonnet 4.5 planning or UI-TARS with GPT-4o planning.
Is Cua free to use?
Yes, Cua is free for local usage with Docker - no account needed. For cloud sandboxes, there is a Free tier at $0 that includes free sandbox minutes and a Linux Small cloud sandbox with full API and SDK access. This is designed for local validation and first cloud tests. Pro plan starts from $10 for active developers with usage-based billing.
What is Cua-Bench?
Cua-Bench is a benchmark suite for evaluating computer-use agents on desktop and mobile tasks. It provides verifiable tasks on real desktop sandboxes with programmatic rewards, oracle solutions, and an RL dataloader. You can run benchmarks interactively, explore tasks manually first, then run with agents and watch progress in real-time with a trace viewer showing screenshots and actions.
How does sandbox snapshot and fork work?
You can define your environment with custom dependencies, snapshot it to save the full sandbox state, and fork into parallel runs instantly. The Snapshot API enables hot-start in under 1 second when restoring from a snapshot. This lets you configure an environment once, save it, then spin up multiple identical sandboxes for parallel agent execution.
What is Lume?
Lume is a CLI for running native macOS VMs on Apple Silicon with near-native speed, delivering up to 97% native CPU speed. It provides local macOS sandboxes with headless mode support and the same Computer SDK as cloud沙boxes. Lume is MIT licensed and open source from day one, launched on Hacker News in March 2025.
Can I integrate Cua with Claude Desktop or Cursor?
Yes, through the MCP Server you can use Cua agents as tools in Claude Desktop, Cursor, or any MCP client. This enables multi-player computer-use for coding agents where agent and human cursors coexist on the same screen. CuaBot launched in February 2026 bringing this multi-player capability to any coding agent.
What are the use cases for Cua?
Primary use cases include AI coding assistants with safe isolated environments for Claude Code and Codex, browser and desktop automation where agents interact with any real UI, cross-platform testing running the same test against Linux/macOS/Windows/Android, and benchmarking agents on OSWorld, ScreenSpot, and custom task sets. Examples include Windows apps behind VPN, PDF to form automation, contact export, and complex UI navigation.