Cua

What is Cua?

Cua is infrastructure for computer-use agents. It scales fleets of virtual computers for parallel training, evaluation and data generation across Linux, Windows, macOS and Android.

Key capabilities of Cua

• Cua Driver: open-source background driver for desktop automation, compatible with MCP and CLI
• Cua Sandbox: local and cloud sandboxes for agents
• Cua Fleets: elastic warm pools of machines on demand
• Cua Bench: framework for authoring evaluations and benchmarks
• Lume: macOS virtual machine management on Apple Silicon
• Multi-OS support: Linux, Windows, macOS and Android environments

How Cua works

Developers provision a sandbox locally or in the cloud, then point a computer-use agent at it to operate the desktop. Fleets keep warm pools of machines so many runs can execute in parallel, and Cua Bench scores agent performance on authored tasks.

Who uses Cua?

Cua is for teams building and training computer-use agents. Its repositories report a community of 27.9K on GitHub, and the vendor lists compatibility with OpenAI, Qwen Code and Factory Droid.

Cua pricing

Cua Fleets is usage-based, at $0.044625 per vCPU per hour and $0.0223125 per GiB of memory per hour. Cua Driver, Cua Bench and Lume are MIT licensed.

Cua alternatives

Alternatives include Scrapybara for hosted agent computers, Kernel for browser infrastructure and Manus for a general agent product.