The Shift from API to OS

The AI agent ecosystem is currently moving past the era of isolated chat interfaces and toward 'computer use'—the ability for models to interact with existing software exactly as humans do. Cua provides the physical and virtual environment where this interaction occurs. Founded in 2025 by Francesco Bonacci, formerly of Xbox's gaming AI division, and based in San Francisco, the company is part of the Y Combinator X25 batch. They have quickly identified a bottleneck in the current AI stack: standard cloud infrastructure like AWS EC2 is not designed for the specific requirements of agents that need screens, mouse control, and rapid state restoration.

Cua offers cloud-based sandboxes that provide more than just a terminal. These environments include a display buffer, a browser, and root access across Linux, Windows, macOS, and Android. This breadth of OS support is a key differentiator, particularly the inclusion of macOS on Apple Silicon, which remains difficult to find in a standard cloud context. By offering high-performance virtual containers that achieve near-native speeds, Cua allows agents to perform complex tasks—like building an iOS app in Xcode or navigating a mobile interface—within a controlled, observable environment.

Infrastructure Built for Speed

One of the primary technical challenges Cua addresses is the latency of provisioning. Traditional virtual machines can take minutes to boot, which is prohibitive for agents that may need to run thousands of parallel simulations for reinforcement learning. Cua utilizes a Snapshot API and a 'hot-start' mechanism that reduces boot times to under one second. This capability is paired with a Fork API, allowing a developer to configure a single environment—installing dependencies, setting up files, and navigating to a specific application state—and then clone that exact state across hundreds of parallel instances for simultaneous agent execution.

To bridge the gap between local development and cloud production, Cua maintains Lume, an open-source (MIT licensed) tool for native macOS sandboxing. This allows engineers to build and debug their agentic workflows locally on Apple Silicon before deploying them to the Cua Cloud. The consistency across these environments is maintained through a unified Computer SDK, which provides a standard interface for screenshots, clicks, typing, and file I/O.

Evaluation and Benchmarking

Beyond raw compute, Cua is moving into the evaluation space with Cua-Bench. Testing agents in 'computer use' scenarios is notoriously difficult because tasks are often non-deterministic. Cua-Bench provides verifiable task environments—such as a replica of a banking app or a Slack workspace—with programmatic reward signals and oracle solutions. This infrastructure is specifically designed for developers working with Claude Code, OpenClaw, and other modern agent frameworks. By recording agent trajectories and providing human-annotated screenshot metadata, Cua facilitates the transition from raw agent execution to systematic policy training and evaluation at scale.