trycua/cua — The Computer-Use Stack Goes Open Source
Anthropic and OpenAI have spent a year talking about computer-use agents. trycua/cua is the open-source version of that whole stack.
It is sitting at 14.3k stars and shipped cua-driver-v0.0.11 on April 26. The repo is not one tool, it is the whole pipeline. Cua Driver runs background automation on macOS without stealing your cursor, so an agent can click and type while you keep working. Cua Sandbox gives the agent a clean OS-level environment with screenshot, mouse, keyboard, and even multi-touch gestures. CuaBot is the desktop CLI that drives a coding agent inside that sandbox. Cua-Bench is the eval framework. Lume manages the macOS and Linux VMs on Apple Silicon underneath all of it. Cross-platform: macOS, Linux, Windows, Android.
The positioning is sharp. Anthropic's Claude computer-use API and OpenAI's Operator both run inside their own walled gardens. cua is what you ship if you want to build on top without paying the inference markup or accepting the rate limits. MIT licensed, 469 releases deep, actively maintained. The driver release on April 26 is specifically about not interfering with the user's cursor while the agent works, which is the exact friction point everyone hits when they try to run a computer-use agent on their own machine instead of a headless cloud VM.
The larger pattern: the agent infra layer is now reproducing the closed APIs of the frontier labs as open source within months. Browser-use did it for browser agents. Beads is doing it for agent memory. cua is doing it for full computer control. If you are building anything that needs an agent to operate a real OS, this is the cleanest starting point that doesn't lock you to a vendor. Repo at github.com/trycua/cua.
← Back to all articles
It is sitting at 14.3k stars and shipped cua-driver-v0.0.11 on April 26. The repo is not one tool, it is the whole pipeline. Cua Driver runs background automation on macOS without stealing your cursor, so an agent can click and type while you keep working. Cua Sandbox gives the agent a clean OS-level environment with screenshot, mouse, keyboard, and even multi-touch gestures. CuaBot is the desktop CLI that drives a coding agent inside that sandbox. Cua-Bench is the eval framework. Lume manages the macOS and Linux VMs on Apple Silicon underneath all of it. Cross-platform: macOS, Linux, Windows, Android.
The positioning is sharp. Anthropic's Claude computer-use API and OpenAI's Operator both run inside their own walled gardens. cua is what you ship if you want to build on top without paying the inference markup or accepting the rate limits. MIT licensed, 469 releases deep, actively maintained. The driver release on April 26 is specifically about not interfering with the user's cursor while the agent works, which is the exact friction point everyone hits when they try to run a computer-use agent on their own machine instead of a headless cloud VM.
The larger pattern: the agent infra layer is now reproducing the closed APIs of the frontier labs as open source within months. Browser-use did it for browser agents. Beads is doing it for agent memory. cua is doing it for full computer control. If you are building anything that needs an agent to operate a real OS, this is the cleanest starting point that doesn't lock you to a vendor. Repo at github.com/trycua/cua.
Comments