Problem

Cua addresses the infrastructure bottleneck in computer-use agents: agents need to see screens, operate native applications, and run inside complete desktop operating systems, but conventional local virtual machines and generic GUI sandboxes are costly and difficult to scale. Cua’s own account identifies 20–40GB storage requirements per VM, incompatibility between Apple Silicon and x86 hardware, and roughly a day of setup and configuration for each new user. The core economic pain is wasted engineering time and idle infrastructure when teams need to run many isolated, repeatable agent environments rather than a single demo.

The principal use case is operating large fleets of computers for agent training, evaluation, reinforcement-learning loops, synthetic data generation, and batch rollouts. Instead of manually provisioning machines for each experiment or workflow, teams can launch disposable environments, reproduce failures from snapshots, and convert agent activity into training data. Practical workloads include parallel web scraping, automated desktop workflows, and testing agents across multiple operating systems.

Product / Service

Cua is an open-source computer-use infrastructure stack paired with hosted cloud services. Its Sandbox runtime provisions disposable GUI environments locally or in Cua Cloud, while Cua Fleets turns those sandboxes into elastic pools of machines that teams claim on demand. The platform supports Linux, Windows, macOS, and Android, with local backends including Docker, QEMU, and Apple’s Virtualization framework; customers can run the stack on their own infrastructure or move to hosted, BYOC, or on-prem deployments.

Cua Cloud is positioned as “Docker for Computer-Use Agents”: it provides preconfigured full-desktop sandboxes accessible through a browser, encrypted VNC, built-in automation dependencies, persistent session storage, automatic cleanup, and pay-per-use billing. Developers use the same Computer and ComputerAgent interfaces locally or in the cloud, can bring their own model-provider keys, and can scale from one to 100 parallel agents in the earlier cloud offering. The benefit is faster deployment, isolation, reproducibility, and cross-OS execution without having to build and operate the underlying VM fleet.

Market

Cua competes in the emerging infrastructure market for computer-use agents, spanning agent sandboxes, desktop automation, model evaluation, and training-data generation. Its closest alternatives are partly adjacent rather than identical: Browserbase provides infrastructure for browser agents, while E2B offers secure computers and isolated sandboxes for AI agents. Cua differentiates through full operating-system support, native desktop interaction, local as well as cloud execution, and a specific emphasis on fleets, benchmarks, evaluation, and reinforcement-learning workloads.

The company is early-stage but has meaningful developer traction. Y Combinator lists Cua as an active Spring 2025 company founded by Francesco Bonacci, with a three-person team, and its open-source repository reported more than 20,000 stars and an MIT license in the research snapshot. Cua has also described working with partners and beta users and offers a live cloud product with pay-per-use positioning. Public sources reviewed do not disclose revenue or a paying-customer count, so Cua should be viewed as an early commercial-stage company with product and community traction rather than one with proven revenue scale; its status as definitively pre-revenue cannot be confirmed.