Mini Machine · No 01 · 2026 Edition
28 chapters · ≈ 2 h read
For the first time, an AI can sit next to you and code, on a small silent PC for €500. No cloud needed, no team needed: you can build whole projects (a site, a tool, an app) right from your desk. This guide is my workbench, shared chapter by chapter so you can build yours.
What we stack, from the metal up
Le monde extérieur
Cloudflare Tunnel · ton domaine · HTTPS
Le réseau privé
Tailscale · SSH · accès depuis le canapé
L'IA locale
Ollama · modèles open-weight · embeddings
L'agent de code
Claude Code · OpenCode · skills · mémoire
Le système
Ubuntu LTS · utilisateur · pare-feu · paquets
Le métal
Mini-PC · CPU · RAM · SSD NVMe
What you'll find
What it won't be
Table of contents
Six chapters, twenty-five chapters of their own. Read in order the first time (each chapter sets up the next), then dip in and out as you need.
A small, silent PC that's on all the time and becomes your AI-assisted development workshop. The goals, the philosophy, and why it changes the way you tinker.
To get you excited: a tour of what you can really build with this machine. Home automation, media server, apps, automations with agents, intelligence offloaded onto your phone.
Which mini-PC, how much RAM, which SSD? The jargon-free buying guide for a machine that runs a coding agent and AI models locally.
Got a Mac mini instead of a Linux mini-PC? A large part of this guide runs on it without changing a thing. Here's what works as-is, and the few settings to adapt.
Why Linux, and how to install it stress-free: from the bootable USB stick to the first boot. We pick Ubuntu LTS, the safe bet.
A handful of one-time gestures: user, automatic updates, smart swap (zram), and the basic tools. The machine becomes clean and stable.
Claude Code or OpenCode, in five minutes. The trick: install it now, so it configures the rest of the machine for you.
Your code deserves a safety net. Git to version everything, GitHub to keep it safe and share it, and a real backup strategy so you never lose anything, even if the machine fries.
Every project in its own watertight bubble. Docker keeps your projects from stepping on each other, contains the damage when the agent experiments, and makes your deployments reproducible. For beginners and experts alike.
An agent isn't a chatbot. It's a model taken out of its box and given tools and a goal. Here's how it really works, no magic.
Give your agent more tools, connect it to your data, delegate to sub-agents, and tighten its permissions. How to go from a default agent to one tailored to you.
Concretely, what is an agent good for in a real project? Code review, research, parallel execution, tasks that run on their own. The patterns that change how you work.
This is where the dedicated machine truly comes into its own: an always-on agent that runs in the background or on a schedule, in private, costing you nothing locally. Your little colleague that never sleeps.
What if your mini-PC became an assistant you drive from your phone? Hermes and OpenClaw turn the machine into a self-hosted agent, reachable by chat, backed by your local models.
Reach your machine from anywhere, a café, a train, your couch, as if it were sitting right next to you. Encrypted, without opening a single port, in five minutes.
Driving the machine from your Mac or your Windows PC: the terminal for everything, and a VNC graphical desktop for when you need one. The right apps, the right config.
Put a service online on your own domain, over HTTPS, without opening a single port on your router. The tunnel goes out from your home, it doesn't come in.
A machine reachable from anywhere, running code generated by an AI: security isn't optional. SSH keys, firewall, secrets, agent permissions. How not to shoot yourself in the foot.
Run real AI models at home, free and private. Installation, your first model, and how to wire it up to your coding agent.
Which local model should you run on YOUR machine? The 2026 open-weight landscape (mostly Chinese), and the simple rule for fitting them to your RAM.
The best of both worlds. Claude Code keeps the lead, it plans, reasons, reviews, and delegates the repetitive heavy lifting to a local model. Private, frugal, and devastatingly effective.
The skill that changes everything. Before writing a single line of code, you turn a fuzzy idea into a crisp spec, by talking it through with an LLM. 80% of the outcome is decided right here.
An agent forgets everything between sessions, except what you write in its memory files. How to structure them so it holds onto the context, your rules, and your decisions.
When you catch yourself repeating the same instructions to the agent, turn them into a reusable skill. The step that takes you from 'assistant' to 'workshop'.
A Claude Code session manager with a real interface: every session, every project, found and resumed in one click. Handy on your working machine, but it's a local tool: we'll see why it's a poor fit for the screenless mini-PC.
The links that matter, the communities where you can ask your questions, and the reflexes for when something jams. Your address book for what comes next.
The agent writes fast, your job becomes checking its work. Best practices for reviewing a diff, auditing quality, and putting security through the wringer, with Claude Code or any agent.
A machine that's always on stays clean over time. The OS, the dependencies, the tools, the agents: what you automate, what you update by hand, and how to avoid breaking anything.
Author's note
The trick isn't the machine. It's how you brief it, and the memory you lend it from one session to the next.
A €500 mini PC runs this stack without breaking a sweat. What separates a finished project from an agent that drifts off is the chapter I spent the most time on: Working well. Framing, memory files, skills, audits. If you only read one chapter, make it that one.
Go to Framing with an LLM →build 2026-06-20 · stack Astro · Tailwind · MDX
No cookie, no tracker, no email capture.