Cross-OS sandbox infrastructure for scaling computer-use AI agents through training, eval, and data-generation workloads.
What it does
Cua provides sandbox infrastructure to run computer-use AI agents at scale across Linux, Windows, macOS, and Android machines, either locally or in the cloud. It supports snapshot-based forking for parallel episodes, warm machine pools for fast batch throughput, and delivers human-verified trajectory datasets for training.
Core features
One API to boot Linux, Windows, macOS, and Android sandboxes
Snapshot-native forking for parallel rollout reproduction
Warm machine pools for low-latency batch claiming
Open-source driver (Cua Driver) with MCP/CLI support
Eval and benchmark authoring via Cua Bench
Human-reviewed, verified agent trajectory datasets for training
Best for
→Running large-scale computer-use agent evaluations
→Generating reinforcement-learning training data from real environments
→Reproducing and debugging agent failures via forked snapshots
→Testing agents across multiple operating systems from one API
Reviews
Big-picture takes: what it's for and whether it delivers. High-engagement YouTube videos — not sponsored.
Tutorials
Step-by-step: exactly how to get things done with it.