Cua

Cua (trycua/cua, MIT, YC X25, ~21–28k stars) is the open-source infrastructure suite for computer-use agents — the 'body' that lets an AI 'brain' like Claude or GPT-4 see a screen, move the mouse, and type across macOS, Windows, Linux, and Android. It bundles five pieces: cua-driver (background desktop control that doesn't steal your cursor or focus, exposed over CLI and MCP), cua-sandbox (ephemeral VMs and cloud desktop fleets via run.cua.ai), Lume (near-native macOS virtualization on Apple Silicon), cua-agent (the agent SDK), and cua-bench (benchmarks on OSWorld/ScreenSpot/Windows Arena plus RL trajectory export).

It is infrastructure, not a finished agent: you still bring the model, the prompts, and the task logic, and cloud fleets are paid capacity pools that need cleanup discipline to avoid ongoing charges. Its differentiator is the benchmark suite — no other open computer-use framework ships evaluation and training-trajectory tooling in the same repo.

Install the driver with the one-line installer (curl https://cua.ai/driver/install.sh on macOS/Linux, PowerShell on Windows); pip install cua for the sandbox SDK. Tell Claude Code, Cursor, or Codex to integrate it via the MCP server.