I run openclaw (and also hermes now to test) in kubernetes pods connected to my matrix instance. They have debug access to the kubernetes api, grafana mcp, and can make PRs to my private forgejo instance to change k8s manifests, which then get deployed with argocd. Oh and I made them build most of that. They run my whole self-hosting setup now.

For everyday stuff they use ollama cloud, agents that handle private information use ollama on my gaming pc.

I have no idea how you could even ensure inference on someone elses hardware would not leek context. But that is a huge issue.