Skip to main content
You rent the model. You own the house. This guide takes you from an empty cloud instance to a locked-down box running a model gateway, a coding agent, a few tools, and a memory that stays yours when the models change. The guide is the reading copy of the own-the-house repository. The scripts, config files and templates that the steps name (such as scripts/lock-edge.sh and gateway/litellm.config.yaml) live in that repository, so clone it before you start.

What you end up with

A single Ubuntu 24.04 ARM box on which:
  • Nothing faces the open internet except one SSH door. Everything else lives on a private Tailscale mesh.
  • One gateway (LiteLLM) fronts every model provider, so switching models is one line of config.
  • An agent reaches the models through that gateway and calls a small set of tools over MCP.
  • A notes vault that the agent reads and writes gives the system a memory, even though the model has none.
The box is free on Oracle Always Free ARM, or a few euros a month on Hetzner. The only real recurring cost is your model subscription.

Run it with an AI agent

The build is designed to be run with an AI assistant as your guide, one verified step at a time. Hand the repository to your assistant and start with AGENTS.md.

The steps

Check each step with scripts/preflight.sh before you move to the next one.

Read before you build

The build stops at the foundation on purpose. Grow into it covers what to add later, and in what order.