What runs where, and who may start it
The question is no longer “ChatGPT, yes or no”. The question is what runs where and who may start it. Local and in-house AI is growing alongside the cloud: a Mac mini handles inference, open weights keep closing the gap, and agents are the product of the year and the risk of the year. This series takes you from your first local model to a control room you run the whole thing from — and to the rules without which it all falls apart.
45 parts · six tracks · more coming
Read it front to back like a course, or jump around by tag. If in doubt, start with the first part of each track — they assume you have installed nothing yet.
Prefer a fixed order to a signpost? Follow the path An agent on your NAS from scratch (13 steps).
01 · 5×
Model, harness, agent — and why the question moved from “AI yes or no” to “what runs where”.
02 · 5×
Two complete setups from hardware to invoice: what runs where, what it costs and why it is put together this way.
03 · 5×
Your first model on your own machine: Mac mini, Windows with an Nvidia card, memory, and when local is enough.
04 · 6×
One gateway in front of every model, API keys in one place, and routing each job to what handles it best.
05 · 8×
A control room in your pocket: Telegram as the panel, a safe route home, and a queue you approve yourself.
06 · 8×
Files from your disk into the chat: how it really works, what RAG is, what MCP does, what stays off the cloud.
07 · 8×
Agents are the product of the year and the risk of the year. One page of rules and the case studies behind them.
Tags
25 parts tagged “AI agents” · Clear filter
Five accounts, three subscriptions, every model in a different tab and the files still sitting on disk. This part is the map of the whole series: what a control room is, which four layers it consists of and in what order to build it.
The model is the engine, the harness is the car, the agent is a driver with a job to do. If you don't tell these three layers apart, you can't make sense of a single news story about AI — or decide what to run at home and what to leave in the cloud.
In August 2026 alone, Grok Bot, Cursor Origin, Slack Code and Claudeforce all came out. They all solve the same thing: letting AI out of the chat and into your work. And they all run into the same question that nobody can answer elegantly — who vouches for it.
Most of this series assumes a box that runs at home around the clock. If you don't have one and don't want one, you can get a large share of the benefit from the app alone. An honest comparison: what works without a server, what doesn't, and how to decide.
An account, a connected GitHub, a first brief — and within a few weeks a website that finds its own topic every morning, writes an article and waits for you to approve it. That's exactly how the site you're reading came about.
Two containers on a home NAS: one holds all your chats, the other is an agent that reads them continuously, writes promises into Notion and answers you in Telegram. Claude prepares the containers; you just log in twice in the terminal.
You don't need a Mac mini to have your own agent. The NAS you keep your photos on is enough: the agent is lightweight, because the thinking is done by a model in the cloud. A half-hour installation and a hundred concrete tasks, grouped by how much you have to entrust to it.
One account, one key, one invoice — and access to models from dozens of providers. What it really costs, how to set a spending cap, when it suits as your main route and when only as a safety net for an agent.
An agent doesn't have one price. It has four components, three of which are predictable while the fourth can spring a surprise. How to measure them, a table to fill in, where the money most often goes missing, and five levers for cutting the bill without losing the benefit.
The difference between a useful agent and a dangerous one is not in the model but in a single line: where preparation ends and action begins. How to draw that line so that it holds — with a folder, a queue and one button on your phone.
Your NAS now runs an agent that reads your messages. That turns a box for photos into a machine worth locking properly. A concrete procedure for DSM: Tailscale, what to switch off, the two scheduled tasks people forget, and a test you can trust.
Ten scheduled tasks mean ten messages you'll stop reading within a week. How to fold them into a single overview that takes a minute to read, what doesn't belong in it, and how to tell it's time to cut it down.
The best ideas arrive when your hands are not free. A voice message in Telegram, a transcript, a reply by voice or text. A four-line setup, a free option and three situations where voice is the wrong tool.
The morning brief did not arrive. The agent does not react to a message. Before you start reinstalling, go through six questions in order from the most common cause to the rarest. With the commands that answer them.
Connectors aren't only for cloud services. A local MCP server can give the model access to a folder, a database or the command line — and the difference between a useful hook-up and a dangerous one is a single parameter: how much you fence off for it.
After three months the agent has your memory, fine-tuned procedures and the keys to everything in a single folder. Which of that to back up, where, how often — and why the backup has to sit somewhere the agent itself can't reach.
Nobody reads the brand manual in PDF, least of all an AI. Put the rules, texts and approved photos on one server and every tool will request them itself — Claude, the agent on your NAS and the colleague in another department. What belongs on such a server and how to build it.
How to build a company knowledge base from the first folder to deployment is covered by the big guide in Tips. This part is a short companion for readers of the series: four rules that tell you a base is sound, trust levels the guide does not cover, and connecting it to an agent on your own machine.
It isn't a bug in the code that someone will fix. It's a property of the way models work — and the only defence is to take one of three capabilities away from the agent. Which three they are, and how to check your own setup.
A twenty-page policy nobody reads protects less than a single page that everyone has on the noticeboard. A complete template with eight points that takes five minutes to read and can be adopted this week.
The plan from January: let agents take over routine work and slim teams down by up to 60%. The reality from August: 220% more code, 36% more finished features, 40% more incidents. A case study you can learn more from than from ten how-to guides.
An agent you have to explain everything to again is just a more expensive chat. A skill is a written-down procedure that it loads by itself next time. How one comes about, what it looks like inside, how to fix it — and why an agent's learning should pass your review just like everything else.
Your agent knows your promises, your mail and your calendar. You can't lend it to the children. A second profile has its own memory, its own bot and its own personality — and this part also says where a profile stops being enough and you need a second container.
An agent for one person is a tool. An agent for five people is a colleague, and it has to be clear who it answers to, who may do what and who pays for it. A three-role model, a shared channel instead of private messages, and five decisions to make before you switch it on.
In the first month you add things. By the third you have twenty tasks, read three and pay for the rest out of habit. A one-hour audit with four questions for every task, a table to fill in, and five patterns that tell you which task won't survive.
Get the next part
New parts and what matters in the world of agents go out in the weekly newsletter.
The Top 30 tips e-book, free — it lands in your inbox right away.
Then 1 tip a week · no spam · unsubscribe in one click