Image: reddit.com · Autonomous AI (Claude) maintenance loop while preventing it from nuking the lab
A daily Claude update loop for a homelab, fenced by an MCP server
n8n runs Claude each morning to update homelab services through a Claude Code-written MCP server that holds all the privileged access, with rollback on failure.
The full story
The lab is a Synology NAS and a 2018 Intel Mac Mini running Debian. Every morning Claude applies updates to the running services in a loop: detect, decide, act, verify, roll back on failure, report and learn. n8n runs the loop and the schedules, and the model is served through Meridian, a Claude Agent SDK proxy in passthrough mode that hands the model's tool calls back to n8n.
Every real action goes through a small MCP server written with Claude Code, the only privileged part, which holds the Docker socket, the compose directory, git and a scoped SSH key to the NAS. Claude can only ask it for specific curated operations; Forgejo stores the config and audit trail, and Prometheus decides whether an update actually worked.
How it was done
Tools
Themes
Loading themes…
Sources
From a short post by the maker, so the details are as they tell them.
- Reddit: Autonomous AI (Claude) maintenance loop while preventing it from nuking the lab
The maker's post on r/homelab.
The poster has deleted it, but the original is still readable (it can take a few seconds to load): archived copy of Reddit: Autonomous AI (Claude) maintenance loop while preventing it from nuking the lab (opens in a new tab)