LLM inference is remote, yes, but tool calls happen locally on the device. So CC knows it has access to modify files, perform bash commands, etc.