trycua/cua ? reverse-engineered prompt

Build me an open source toolkit called Cua for people building AI agents that can use computers.

I want agents to be able to safely control full desktops in sandboxes. They should see the screen, click, type, take screenshots, run shell commands, use the clipboard, and complete normal app or browser tasks without taking over my real cursor. It should work across macOS, Windows, Linux, and Android where possible, with local and cloud sandbox options.

Please include a friendly command line setup, a Python SDK with simple examples, and a way to run common coding agents inside the sandbox. I also want a benchmark tool so people can test agents on desktop tasks and export the results for training. For Apple Silicon, include a simple way to create and run fast macOS or Linux virtual machines.

Make the repo feel polished, with clear docs, install scripts, examples, tests, and a quick start that gets someone from zero to a working agent demo fast. Look up current docs online if you need to.