The AI stack on Proxmox.

An LXC container is a lightweight virtual Linux machine — much lighter than a full VM. You'll make one container, called something like ai-box, and everything else in this guide runs inside it.

Why this first: one always-on container is the home for all your AI tools. Build it once and it quietly runs 24/7, ready for every device on your network.

In a browser on any device, go to your Proxmox box (replace with its IP address):

https://YOUR-PROXMOX-IP:8006

Open the host Shell (Datacenter → your node → Shell) and download an Ubuntu template:

pveam update
pveam available --section system
pveam download local ubuntu-24.04-standard_24.04-2_amd64.tar.zst

The exact filename from the available list may differ — use the ubuntu-24.04-standard one shown in your list.

Click Create CT (top-right). Set a hostname (e.g. ai-box), pick the template you just downloaded, give it about 4 GB RAM and 2 cores to start, finish the wizard, then Start it.

The CTID is the number shown for your container (e.g. 100):

pct enter CTID

The container is a small Linux machine, so from here the setup matches the Linux guide. First the basics and Node.js (the engine three of the AI tools run on), then Claude Code — Anthropic's AI coding agent that lives in your terminal and edits real files for you.

Why this first: every tool below installs through the terminal, and three of them need Node. Get this done and the rest is copy-paste.

apt update && apt upgrade -y
apt install -y curl git build-essential

This adds the official NodeSource repository, then installs the current LTS version:

curl -fsSL https://deb.nodesource.com/setup_lts.x | bash -
apt install -y nodejs

npm install -g @anthropic-ai/claude-code

Go into any folder you want to work in, then launch:

claude

The container has no desktop browser, so the first run prints a URL. Copy that link, open it in the browser on your laptop or phone, sign in to your Anthropic account, and you're authorized. After that, just type what you want in plain English.

OpenAI's command-line coding agent. Same idea as Claude Code, different brain — handy as a second opinion or when you've used up one tool's quota.

Why you want it: variety. Different models are stronger at different things; having both means you're never stuck.

npm install -g @openai/codex

codex

First run lets you sign in with your ChatGPT/OpenAI account (or an API key). With no desktop browser it prints a link to copy into your own browser, same as Claude.

Google's command-line AI agent. Notable for a very generous free tier and an enormous memory (context window) for long documents.

Why you want it: free headroom. Great for big, sprawling tasks before you spend on the others.

npm install -g @google/gemini-cli

gemini

First run signs you in with a Google account using the same URL-based login — it prints a link, you open it in your browser to authorize. The free tier is large, so you likely won't pay anything to start.

A coding agent from Nous Research — "the agent that grows with you." Same idea as Claude Code, Codex, and Gemini, with two twists: it's self-improving (it learns your preferences over time), and it's model-agnostic — you point it at whichever brain you want (Nous, OpenAI, Anthropic, OpenRouter, and more).

Why you'd add it: it's a fourth tool in the rotation and the most model-flexible of the bunch — handy for staying un-locked-in (see the future-proofing principles).

You install it inside the Linux container you made in Step 1 — same as the other agents on this page. It installs natively in the container.

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

pip install hermes-agent
hermes postinstall

hermes setup        # configure it the first time
hermes --tui        # start it (modern terminal UI, recommended)

During setup it asks how you want to sign in — an API key, or an OAuth login via hermes setup --portal. It works with the Nous Portal, OpenAI, Anthropic, OpenRouter, and others, so you can reuse an account you already have.

A private network (a "mesh VPN") that connects all your devices to each other securely — your Proxmox box, your PCs, your Macs, your phone, your NAS — as if they were in the same room, from anywhere in the world.

Why you want it: this is the glue. Once your AI box is on Tailscale, every device on your network can reach it by name — to share one Ollama for the whole house, or to use Claude Code on it from your phone.

You can run Tailscale on the Proxmox host itself:

curl -fsSL https://tailscale.com/install.sh | sh
tailscale up

An LXC container needs access to the host's TUN device first. From the Proxmox host, edit /etc/pve/lxc/CTID.conf and add these two lines, then restart the container:

lxc.cgroup2.devices.allow: c 10:200 rwm
lxc.mount.entry: /dev/net/tun dev/net/tun none bind,create=file

Then, inside the container, install and bring it up:

curl -fsSL https://tailscale.com/install.sh | sh
tailscale up

This prints a URL. Open it in your browser and sign in with the same Tailscale account you use on your other devices, so they all join one private network.

Runs AI models on your own box instead of the cloud — free, private, and works offline. Good for chat, summarizing, and coding help without a subscription.

Why you want it: no per-use cost, nothing leaves your machine, and combined with Tailscale (Step 5) this one always-on box can serve models to your phone and every other device.

Inside the container, install Ollama and pull a small, capable model:

curl -fsSL https://ollama.com/install.sh | sh
ollama run llama3.2

The first run downloads the model (a couple of GB), then you can chat right in the terminal. Type /bye to leave.