Home
Local workspace on Ubuntu 24.04
cua-sandbox
eval-runs
screenshots
datasets
README.md2 KB
run_eval.sh4 KB
episode-042.json18 KB
Scale computer fleets
for every
Run training, eval, and data-generation workloads in parallel on Linux, Windows, macOS, and Android machines. Fork from snapshots, reproduce failures, and turn agent activity into training data.
Why Cua
Where generic sandboxes stop.
FLEET COVERAGE
Run one rollout layer across four OS families.
LINUXLinux for code-heavy rollouts
Ubuntu and custom images for browser, terminal, and code-heavy workloads.
WINDOWSWindows for enterprise UI
Windows 11 environments for enterprise apps, desktop UI, installers, and file workflows.
MACOSmacOS on Apple Silicon
Bare-metal Apple VZ macOS VMs for Finder, Safari, permissions, and native app flows.
ANDROIDAndroid mobile state
Android emulation for touch, swipe, auth flows, mobile app state, and notifications.
Products
From a single sandbox to an infinite fleet.
Cua Sandbox
The computer-use runtime: boot a cloud or local VM across Linux, Windows, macOS, and Android from the Python or TypeScript SDK, or the cua CLI.
Cua Run
Elastic infrastructure that scales sandboxes into warm pools of machines you claim on demand for evals, RL loops, data generation, and batch rollouts.
SOC 2-readyBYOC availableMIT licensedOn-prem available
Open-source control and eval layers: Cua Driver · Cua Bench · Lume
Open Core
It starts with Cua Driver, the open-source computer-use tool behind Clicky.
Cua Driver
The open-source background computer-use driver for native desktop apps. It lets agents click, type, scroll, inspect accessibility trees, and capture window state through the same MCP/CLI surface without stealing your cursor or focus.
Open-source and MIT licensed. One binary can run as an MCP stdio server, long-running daemon, or one-shot shell command on macOS and Windows, with Linux in pre-release.
macOS + Windows
Native desktop backends preserve the user session; Linux support is in pre-release.
Drops into your stack
Run the same binary as an MCP stdio server, long-running daemon, or one-shot shell command.
MCP server + CLI
Exposes an MCP server and a CLI, so agents and humans drive the same harness.
Cua Sandbox
Local and cloud sandbox provisioning for agent training and evaluation.
Cua Bench
Eval and gym authoring for cross-surface agent benchmarks.
Lume
macOS VM management on Apple Silicon, built on Virtualization.framework.
SOC 2-readyBYOC availableMIT licensedOn-prem available
Verified Data
Train against the environments, or take the verified data they produce.
Two ways to consume the same fleet, and either way the hard part is verifying the data.
Train against live environments
Point your training or eval loop at warm pools of real machines. Claim an environment on demand, step through the task, collect reward, and release it back. The pools are sized for large parallel batches.
Take the verified data they produce
Skip the harness entirely. We run the rollouts on the same environments and deliver verified trajectory datasets, packaged for your data-ingestion pipeline and scoped by task and surface.
Evaluators in the loop
Every trajectory is scored against an evaluator for the task. A rollout counts only when it completes the objective.
Human-reviewed golden trajectories
Accepted runs pass human QA before delivery. Golden trajectories carry step-level annotations, so you train on verified steps.
Acceptance you can set
Define the bar (target model, difficulty tier, surface mix) and we hold delivery to it before the data reaches your pipeline.
4 OS families
Linux, Windows, macOS, Android
1 driver surface
MCP + CLI across machines
17.8K
GitHub humans
Pricing
Scale computer fleets on your own infrastructure.
Start with the open-source stack. Move into hosted, BYOC, or on-prem infrastructure as concurrency, OS mix, and compliance needs grow.
Free tier on GitHub. Dedicated fleets by request.
