I run an OpenClaw VM and used Claude Code to build the VM scripts. The VM is connected to local llama.cpp, so OpenClaw and the models are running on my own physical hardware.