Sort of, they have no "hands", LLMs can only respond that they want to execute a tool/command. So they do that a lot to: read files, search for things, compile projects, run tests, run other arbitrary commands, fetch stuff from the internet etc.

Obviously the LLM inference is super heavy, but the actual work / task at hand is being executed on the device.

The AI part yes. But they also use quite inefficient rendering on the cli.

yeah I run Claude code on a 2013 Mac book air that refuses to die, I don't think it's very compute heavy.