In one day, this homelab went from manually administered to conversationally operated. Here is how — and what broke along the way.
Five interfaces to everything
The lab now exposes five MCP servers to its AI operator: both Docker engines, the Node-RED instance, the Proxmox hypervisor and this WordPress site. That last hop mattered most — hypervisor access turned the operator from a container janitor into an infrastructure engineer.
The 15:00 incident
Mid-afternoon, every service on the container host stopped responding. Pings were fine, TCP ports accepted connections, but nothing answered. Probes from the second node showed the signature of an I/O stall: load average 136, static pages taking 17 seconds.
Once the hypervisor came online in MCP, the real cause surfaced in seconds: the Docker host was an LXC container with just 4 cores and 8 GB of RAM — carrying 31 containers including OpenProject, WordPress and MariaDB. The morning reboot had been the container itself cycling under pressure.
The fix, at the right layer
One resize call later — 6 cores, 16 GB — and a controlled reboot, all 31 containers self-healed through restart labels. The same session also caught a neighbouring Node-RED container burning 99% CPU in a looping flow, and reclaimed 24 GB of orphaned Docker images across both nodes.
Diagnosis from kernel symptoms, fix at the hypervisor, verification inside the containers — one conversation, three layers.
The site you are reading was published the same evening, by the same operator, through the same protocol.