Cua is an open-source platform designed to scale Linux, Windows, macOS, and Android computer fleets for computer-use agents. It provides a comprehensive ecosystem for agent training, evaluation, and data generation by enabling parallel workloads across diverse operating systems. Created to bridge the gap where generic GUI sandboxes fall short, Cua offers background drivers, cross-OS fleet management, and batch throughput capabilities. Its core components include the Cua Driver for background input, Cua Sandbox for isolated environment provisioning, Cua Fleets for elastic infrastructure scaling, and Cua-Bench for quantifying agent mastery.

Functionality centers on providing a reliable interface for AI agents to interact with computer environments. By allowing agents to send clicks and keystrokes to specific windows without requiring the system cursor or forcing focus, Cua enables agents to work in the background alongside human users. It supports a wide range of operating systems, including Linux, Windows, macOS, and Android, and utilizes a standardized Model Context Protocol (MCP) and command-line interface to facilitate seamless integration into existing developer workflows.

Some of the key features are:

• Background Input: Agents can interact with desktop windows without moving the system cursor or interrupting the foreground workspace.
• Cross-OS Fleets: A unified API allows for the deployment and scaling of Linux, Windows, macOS, and Android machines via local runtimes or Cua Cloud.
• Batch Throughput: Capability to claim pre-booted machines in milliseconds from verified warm pools that scale to zero when idle.
• Cua Driver: An open-source, MIT-licensed binary acting as an MCP server, daemon, or one-shot CLI tool for cross-platform desktop automation.
• Cua-Bench: A specialized collection of desktop and mobile tasks designed to evaluate and train agents, providing verifiable metrics for agent performance.
• Native macOS Virtualization: A proprietary, patented virtualization stack specifically engineered to scale macOS environments for cloud-based automation.
• Extensive Environment Catalog: Access to thousands of environments and tasks spanning productivity, creative, developer, and scientific workflows.

Operationally, Cua enables users to either point their training or evaluation loops at warm pools of real machines on demand or consume pre-collected, verified trajectory data. The platform provides granular control over environment state, allowing for repeatable runs, snapshot-based forking, and objective-based evaluation through task-specific integrity checks. It integrates with existing agent frameworks through the MCP protocol, ensuring that the same harness is used for both development and production.

Some common use cases include:

• Agent Benchmarking: Using Cua-Bench to measure agent performance across expert-authored tasks like KiCad schematic design.
• Automated Data Generation: Scaling agents to perform complex, multi-step workflows to collect verified human-quality trajectory data for training models.
• Parallel Testing: Running GUI-based automated tests or evals across thousands of instances to accelerate the agent development lifecycle.
• Background Automation: Deploying agents to perform repetitive desktop tasks—such as file manipulation, report generation, or CAD editing—without blocking the local machine's active user session.