cua

Product Introduction

1. Definition: Cua is a specialized infrastructure-as-a-service (IaaS) and runtime platform designed for scalable computer-use agent development. It provides a unified API and toolset for provisioning, managing, and interacting with disposable GUI environments across multiple operating systems.
2. Core Value Proposition: Cua exists to solve the critical infrastructure bottleneck in training, evaluating, and generating data for AI agents that interact with graphical user interfaces (GUIs). It enables AI researchers and engineers to run parallel GUI automation workloads at scale across Linux, Windows, macOS, and Android with reproducibility and verified data output.

Main Features

1. Cross-OS Fleet Management: Cua provides a single control plane to boot and manage heterogeneous fleets of virtual machines and containers. It supports Ubuntu Linux, Windows 11/Server 2025, macOS Sequoia/Tahoe via its patented Cloud macOS stack, and Android VMs. This allows for consistent agent rollout testing across different desktop and mobile surfaces from one codebase.
2. Cua Driver (Background Computer-Use Driver): An open-source, MIT-licensed binary that enables agents to interact with native desktop applications without stealing system focus or moving the physical cursor. It operates as a background service, exposing a standardized Model Context Protocol (MCP) server and CLI interface for clicking, typing, scrolling, and inspecting application accessibility trees on macOS, Windows, and Linux.
3. Warm Pool Infrastructure & Batch Throughput: Cua Fleets maintain pools of pre-booted, "warm" machines that can be claimed on-demand in milliseconds for high-concurrency workloads like reinforcement learning loops or large-scale evaluations. The pools are formally verified for state correctness and can scale to zero when idle to optimize costs, providing high batch throughput for parallel agent tasks.
4. Integrated Evaluation & Data Generation Suite: The platform includes Cua Bench for authoring and running cross-surface agent benchmarks, and Cua Sandbox for local/cloud sandbox provisioning. It facilitates a "eval-in-the-loop" workflow where tasks can include automated evaluators. Crucially, it offers the option to consume either live environments for training or to receive verified trajectory datasets generated from rollouts run on its infrastructure, complete with human-reviewed golden trajectories and step-level annotations.

Problems Solved

1. Pain Point: The extreme difficulty and resource intensity of scaling GUI-based agent training and evaluation beyond a single machine. Traditional methods involve manually managing VMs, dealing with OS-specific automation tools, and lacking reproducibility, which hinders rapid iteration and reliable benchmarking.
2. Target Audience: Primarily AI researchers and machine learning engineers focused on building "computer-use" or "agentic" AI. This includes teams at AI labs developing frontier GUI agents, startups building desktop automation co-pilots, and enterprises creating internal workflow automation agents that require interaction with real software like CAD tools, IDEs, or enterprise resource planning systems.
3. Use Cases: Large-scale parallel evaluation of a coding agent across hundreds of Linux containers running VS Code; training a customer service automation agent on a fleet of Windows VMs with legacy CRM software; generating a verified dataset of mobile app interaction trajectories from Android emulators for imitation learning; reproducing a specific agent failure on an exact snapshot of a macOS environment.

Unique Advantages

1. Differentiation: Unlike generic cloud VMs or simple GUI sandboxes, Cua is purpose-built for agent workflows with features like background drivers, environment snapshots for reproducibility, and integrated eval tooling. Compared to building in-house, it eliminates the massive engineering overhead of maintaining cross-OS virtualization stacks, warm pool orchestration, and data verification pipelines.
2. Key Innovation: The combination of the open-source Cua Driver (providing a standardized, non-intrusive control layer for native apps) with the Cloud macOS virtualization stack (enabling scalable macOS fleets on Apple Silicon, a historically locked-down platform) creates a uniquely comprehensive surface coverage for agent development that is not matched by other platforms.

Frequently Asked Questions (FAQ)

1. What is computer-use AI and how does Cua support it? Computer-use AI refers to artificial intelligence agents designed to operate computers by interacting with graphical user interfaces, similar to a human user. Cua provides the essential infrastructure to develop these agents at scale, offering the real OS environments, background control drivers, and high-throughput fleet management needed for training and evaluation.
2. How does Cua handle macOS automation given Apple's restrictions? Cua has developed a patented virtualization stack, Cloud macOS, built on Apple's Virtualization.framework. This allows them to legally provision and scale macOS VMs (Sequoia, Tahoe) on Apple Silicon hardware in their cloud, providing a crucial surface for agents that most cloud providers cannot offer.
3. Can I use Cua locally or do I have to use the cloud? Yes, Cua supports hybrid deployment. The Cua Sandbox runtime and Cua Driver are open-source and can run locally using backends like Docker, QEMU, and Apple VZ for development and small-scale testing. Cua Fleet cloud service is used when you require elastic scaling, warm pools, and higher concurrency.
4. What does "verified data" mean in the context of Cua? Cua can deliver "verified data" by running your agent tasks on its managed fleets and applying predefined evaluators (automated or human) to each rollout. The resulting trajectory datasets are certified to have been generated in real, unmodified environments and to meet specific acceptance criteria (e.g., task success rate), ensuring high-quality data for training or benchmarking.
5. Is Cua suitable for mobile app automation and testing? Yes, Cua supports Android VM fleets, allowing for the installation of APKs and automation of taps, swipes, and multi-touch gestures. This makes it applicable for developing and evaluating AI agents that operate mobile apps, as well as for large-scale mobile UI testing.