In short: in the space of a month, Meta, OpenAI and Anthropic all put their personal AI agents on computers of their own, in the cloud — and asked for your email and your logins to go with them, and in Muse's case your card. I have been building the opposite — Skynet, a personal AI agent for your Mac, where every agent is a folder on your own machine. Today I run 42 of them. Here is why I went this way, and what it costs.
What changed on 6 October
Anthropic's own help page says it plainly: "new Cowork tasks on Pro and Max plans run in the cloud", the Only on your computer setting is gone, and your sessions and files "live with your Claude account". A cloud session can reach the folders on your Mac "only while the desktop app is open". (Anthropic)
I understand why. A cloud session keeps going when the lid is closed. But reading it, I recognised a fork in the road I had already taken myself — and I went the other way.
It isn't only Anthropic
Anthropic wasn't first. A month earlier, on 8 September, Meta launched Muse: an agent that "runs on its own dedicated computer in the cloud", is "free for most of what people need", reads your email (and can send it, if you let it), and pays for things with your card through Stripe. (Meta) On 29 September OpenAI launched dots: agents with "their own cloud computer" and browser, connected to your apps, signing in with your saved passwords, and searching through your apps in the background for ways to help. (OpenAI)
They are impressive, and they are careful: Meta pays with a one-time card so your real card details stay hidden, and OpenAI says the model never sees your passwords. But look at what you hand over. Your inbox. Your logins. A way to pay. Everything you ask it, and everything it learns about you. All of it lives on a computer that belongs to someone else.
And one of them is free. Meta says Muse data doesn't go to its ad systems, and you can opt out of it training their models. Still, Meta is an advertising company, and when a service is free, it is worth asking what pays for it. Giving the most complete picture of your life anyone has ever had to the company that sells ads against your attention is not something I would do.
How I got here
I have built software for more than 25 years, since I co-founded itslearning in 1999. In 2015 I co-founded Quantfolio, where we build investment advice software — robo-advisors among it — for some of the largest financial institutions in the Nordics. Software that makes decisions for people is not new to me.
I was in GitHub Copilot's technical preview in November 2021, and signed up with OpenAI in May 2022 — half a year before ChatGPT made everyone else notice. It was clear early on what it was missing: it could talk, but it couldn't do anything. So I started giving it hands.
I was building coding agents — AI that writes code, runs it, reads the errors and tries again — before there was a word for them, and building on Claude's API in early 2024, a year before Anthropic shipped Claude Code.
And I used everyone else's. Copilot, Pythagora, Cursor, Windsurf, Claude Code — I paid for most of the tools that promised to let AI do the work, and used them hard. Each was good at its slice. None of them was what I wanted: one place where my agents live, remember, work on a schedule and use all my tools. I kept waiting for the feature that would tie it together, and it never came. So I stopped waiting and built my own.
By April this year my own assistant, Skynet, ran on a server at home with more than a hundred tools — calendar, finance, shopping, LinkedIn — and it posted its own hello to LinkedIn. By May each agent had a computer of its own: a sandbox where it could install things, write and run code, use a browser and a terminal, and keep going for hours. That was three months before xAI launched Grok Bot as an agent with a cloud computer of its own. In August Skynet got a real desktop you could watch live and step into, for the sites that want a human to log in.
And then I moved it. I work on my laptop. My programs are there, my files are there, and so is the browser I am logged in to everything with. The computer in the cloud was always a poor second: an empty machine, far from my files, with a browser that the web treats as a bot — logins, captchas, two-factor checks and bot detection at every turn. So on 18 August I rebuilt Skynet as a plain Mac app that runs right where my work is. It has shipped 64 releases since, and the server at home is switched off.
The cloud agents are learning the same lesson. OpenAI lets you give a dot permission to connect to and use your laptop. Grok Bot's desktop app can run commands on your own computer. Cowork reaches your folders through its desktop app. They all end up reaching back to the machine on your desk — because that is where the work is.
What a day looks like now
I hardly type. I dictate into Skynet, and an agent does the work. Each agent is a folder: its instructions, its memory and everything it has made, as plain files. There is one for my accounting, one for legal questions, one for design, one that writes code. They reach Slack, the books and the rest over their APIs, through 133 small skills — many of them written by the agents themselves, when I asked for something new.
I open the other apps rarely, and mostly to review what was done. Some agents run on a schedule — a morning briefing, a weekly review. When I'm away from the desk, I write to them from Telegram on my phone, and the work runs on my Mac.
Why my agents live on my Mac
| Cowork, Pro and Max, since 6 Oct | Skynet | |
|---|---|---|
| Where the work runs | Anthropic's servers | your Mac |
| Where your files and sessions live | with your Claude account | in folders on your Mac |
| Which AI | Claude | Claude, ChatGPT's codex, or others through pi — per agent |
| Works with the Mac asleep | yes | no — the Mac must be awake |
| What it costs | your Claude plan | your AI plan, plus €99 once for Skynet |
Three things decided it for me. My files stay where they are, and the only thing that leaves is what a task sends to the AI I picked. An agent is a folder of text, so when a better model comes along I switch, and nothing is lost. And it runs on the subscriptions I already pay for, not on a per-token bill.
The price is the second-to-last row: the Mac has to be awake. A Mac mini on the desk, or a laptop on its charger set not to sleep, does the night shift. For me, that is a fair trade.
Cloud inference is temporary
To be honest about it: today, most of the thinking still happens in the cloud. When an agent works, what the task needs goes to Claude or ChatGPT and the answer comes back. That is the part of the stack that isn't ready to come home yet.
But it is coming. Every year the models get smaller for the same quality, and every year the laptops get more powerful. The day a Mac runs a model good enough for real work, I want everything else to already be there: the agents, their memory, their files, their schedules, all on the machine you own. Skynet can already run an agent on a local model today, through pi and Ollama. Moving an agent from the cloud to your own Mac is a setting, not a rebuild — its folder, memory and files stay exactly as they are.
So I am not building Skynet for this year's models. I am building it for the computer you will have in a few years, when the cloud is optional.
That is where I draw the line. Plenty of AI belongs in the cloud: the systems a company runs together, its data and its customers, are built to live there, and should. But a personal agent is different. It reads your mail, knows your calendar, holds your files and remembers your life. That one should live on a computer you own.
Try it
Skynet is free for 30 days on every Mac, with everything in it and no card. After that it is €99 once, VAT included, with 24 months of updates — no subscription to me. You already pay one for the AI.
Download Skynet — 30 days free on every Mac, then €99 once. See Pricing.