Meet the AI agent that runs our homelab: Claude Code in an LXC container

Why a dedicated container

An LXC container is cheap, isolated and easy to snapshot, so the agent has its own place to work without touching the apps it tests. It has its own SSH key to the Proxmox host, which lets it run container commands directly instead of asking a human to copy and paste output.

If that access is ever lost, for example after a rebuild, we generate a new key and add it again. Nothing about it is permanent magic.

What it actually does

• Installs apps in LXC containers, tests them and records real RAM use, which is where the numbers in our catalog come from.
• Builds and maintains this site (Next.js, TypeScript, Tailwind, two languages), runs lint and build, then commits and pushes to GitHub so Vercel deploys it.
• Runs GPU workloads in other containers: image generation with ComfyUI, and training a voice model with Applio, one of the apps in our directory.
• Writes the copy for our weekly notes and guides, and schedules Facebook posts through the Graph API.
• Debugs the network, for example a router that let devices bypass our DNS filter, which we fixed by changing which DNS addresses the router hands out.
• Keeps its own notes in small memory files, so what it learned in one session is available in the next.

Guardrails

• Anything outward-facing, such as publishing, scheduling a post or deleting something, is shown to the owner first and happens only after approval.
• Site content is add-only: nothing is removed from the site or the Facebook page without explicit consent.
• Access tokens are pasted for one task, used from a temporary file and deleted afterwards. They are never stored in the project.
• Claude Code's permission system can block risky commands. When one is blocked the agent says so and never disguises the command to get around it.
• Every change goes through lint and build before it is pushed.

Where it fails

• A heavy image-upscaling test once used all the RAM on the Proxmox host and hung it until it was rebooted by hand. Since then heavy tests run inside containers with a memory cap.
• It cannot touch the hardware, and it cannot create accounts. Creating a Facebook page, signing up on dev.to or Reddit and approving access are done by a human.
• It makes mistakes: a favicon file it generated was invalid and broke the build, and an image description in our catalog does not match its screenshot. Verification caught the first, and a human reads the rest.
• Its numbers and claims need checking. We label RAM figures as estimates for that reason.

How to set up your own

1. Create a small LXC container on your Proxmox host just for the agent.
2. Install Claude Code by following Anthropic's official instructions.
3. Generate an SSH key in the container and add its public key to the Proxmox host only if you want it to manage containers. Start with read-only tasks.
4. Write a rules file (CLAUDE.md) with your boundaries: what needs approval and what must never be deleted.
5. Let it keep notes, and review its work before anything goes public.