Use it: drop ezos.py next to your script, then import ezos. Every action returns {"ok": True/False, "message": ...} so your AI always knows if it worked — and on failure you now get a visible WARNING ezos: .... Both ezos.open_app(...) and ezos.openApp(...) work. For a Jarvis, let your LLM call ezos.run("open_app", app_name="spotify") and show it its toolbox with ezos.describe_commands().

New — give the AI its own cursor 👁️: ezos.give_cursor() + ezos.give_keyboard() hand a vision model a screenshot with a labelled coordinate grid (grid_screenshot()) so it can see the screen, move its own mouse and type. Wire it all together with ezos.VisionAgent(OpenAI()) and just say "play karan aujla on spotify".