n8n Now Runs My ENTIRE Homelab

A home lab can be run like an always-on IT desk by pairing n8n with an AI agent (“Terry”) that monitors services, troubleshoots failures, and—after explicit approval—executes fixes across Docker, servers, and network tools. The core idea is practical: start with tightly scoped permissions, teach the agent repeatable troubleshooting steps using real CLI/API tools, then expand capability only when guardrails and human-in-the-loop approvals are in place.

The build begins with “Baby Terry” and upgrades to a version that can do more than check whether a website responds. Terry is given access to concrete tools: an HTTP request tool to verify a site is up, and an SSH-based command runner (implemented via an n8n SSH node converted into a subworkflow) to inspect the host and manage Docker containers. The workflow is structured around a simple loop: Terry checks the website, and if it’s down, he runs the same commands a human would—first confirming container status with docker ps, then using docker inspect and docker logs to identify why the container failed. Early tests show Terry can detect a stopped container and report exit details, and the agent improves when prompts explicitly require log review.

From there, Terry shifts from “chat-driven” to scheduled operations. A schedule trigger runs every five minutes, but the workflow must be adapted because the agent’s memory and user prompt were originally tied to chat sessions. The solution is to inject a prompt and a chat/session ID via set-field nodes, then route Terry’s results to Telegram for notifications. To avoid noisy alerts, Terry is forced into structured output (JSON fields like websiteUp boolean and message text). An if node filters outcomes so Telegram messages are sent only when the website is down.

The next leap is repair, not just diagnosis. Terry’s prompt is modified so that when the website is down, he attempts a docker start and then re-checks the site. A more challenging test introduces a port conflict by running a Python server on the same port as the Dockerized site. Terry initially fails in a controlled way—he sticks to known playbooks—then a “more powerful” prompt allows broader CLI troubleshooting. That expansion reveals a key risk: the agent can take destructive actions (even stopping the wrong process), which leads to the video’s main safety mechanism.

Human-in-the-loop approval is added using an n8n “human in the loop” Telegram step. Terry must request permission before commands that modify the system. The agent outputs structured fields such as needsApproval and commandsRequested; the workflow sends an approval request to Telegram, and only after approval does it loop the approved instruction back into Terry. With this guardrail, Terry can fix the port-conflict scenario by identifying the conflicting Python process, terminating it, restarting the Docker container, and confirming the website is operational—while the human remains in control.

Finally, Terry is promoted beyond the sandbox: secure remote access is enabled via Twingate so the cloud-hosted n8n instance can reach the home lab. Terry is then reconfigured with new “personas” and tools for UniFi (API-based bandwidth analysis), Proxmox (API/CLI for VM inventory), and Plex (API for active streams and control). The takeaway is less about one perfect workflow and more about a repeatable pattern: connect monitoring to an agent, teach troubleshooting with real commands, constrain actions with structured outputs and approvals, and scale to multiple systems as documentation, sub-agents, and help-desk workflows are added in future steps.