I was running it with my £20 / month OpenAI plus Codex account until I started using Codex to code, so I switched to a $20 / month Ollama subscription model, it primarily uses a Minimax model.
It can also switch to my own Fedora inference box which runs a Nemotron model quite well without a GPU.
My OpenClaw runs on a dedicated RaspberryPi, which it has full control of. I gave it full access to my webserver mikehardcastle.com and has so far behaved itself very well. I stress through prompting how important this server is, so it takes great care to manage it properly.
