We're in the Homebrew Computer Club days of personal AI
TL;DR: Your data is scattered across dozens of companies' servers, each with its own half-built AI. I think the next big shift is bringing it home: a box on your home network that holds all of your data, runs your own AI, and only calls the big AI labs when a problem is worth it. The pieces already exist, but only for tinkerers. We're in the Homebrew Computer Club days of personal AI, and nobody has built the Apple II yet.
A hardware residency in Shenzhen just opened, looking for founders building health and longevity devices. I replied asking someone to build an AI-first Whoop replacement that isn't exorbitantly expensive. Someone from their team replied: you should build it.
Maybe. But I think there's a much bigger opportunity here: bringing the data from every device you own back home, and bringing the intelligence home with it.
Locked in everywhere
Let's start with my Whoop. I love it. It's one of the few products that has genuinely changed my life. But all of my data lives on Whoop's servers, the build quality is poor
, and it's expensive: $199 to $359 a year in the US, and about 50% more in Thailand, where it has become something of a status symbol. Amazfit sells a band with similar sensors for $99, once, with no subscription. That's half of one year of Whoop's cheapest plan. I should be able to swap a fitness band as easily as I swap a t-shirt.
To be fair, Whoop isn't the worst offender here. It can share my sleep, recovery and strain with Apple Health, and I can download an export. But that just moves my data from one company's walled garden to another's, or gives me a snapshot of it. I still don't own it. The software is the part that's easiest to replace now: with AI, I could rebuild the parts of the app I actually use, and make them my own. What I really want is to own the data myself.
And it doesn't stop at my wrist. I recently tried to pull my own financial data together. Interactive Brokers makes it easy: they even run their own MCP server, so shout out to them. With Chase, to get my own transactions into Moolot, the product I'm building, I had to go through Plaid, a paid third party, and Plaid has to approve my app before I can even do that. Meanwhile, Chase now charges aggregators like Plaid to pass along its customers' own data. For now, it's easier to have my agent open Chase in its own browser and read my transactions for me.
It's my money and my record of it, and I still have to ask permission to see it in my own tools. Do the same exercise with your photos, your messages, your health records and your playlists. Each one lives in someone else's database, usually with its own half-built AI bolted on top, and general-purpose software that isn't tailored to you. Your digital life is scattered across the internet, and none of it is really yours.
Things are moving toward connectors anyway. Interactive Brokers already offers one, and more companies are adding them every month, so your data can flow straight into AI. That's great, but it raises the real question: when all of it flows somewhere, where should that be? Do you want your bank, your health, your messages and your plans all streaming into Anthropic, OpenAI or whichever AI company you use this year? Or do you want that data to land on a server you own, and let the AI come to it?
Software built around you
Owning your data isn't only about privacy or switching. It makes software better, because it can finally be built around you instead of the average person.
Language learning is where I feel this most. I'm intermediate in Chinese, Japanese and Spanish, and every time, the hardest part was the jump from the textbook to real life. Courses are built for the average learner. When I first got to Taiwan, learning to order a coffee would have been far more useful than the words for eraser and dorm room, and my textbook didn't teach me to ask where the bathroom was until chapter 15.
That's why I'm building Shangri Lingua, to teach me from the language I actually use every day. When your data lives with you, every app can work this way, from your language to your money to your health.
Protecting your attention
I saw a video of a building contractor who starts every morning the same way: his AI agent prints his plan for the day on a sheet of paper. He doesn't open his phone first thing, so his mind doesn't get hijacked before the day starts. The agent did the screen time for him.
That's the opposite of how most of our tech works. Social apps are built to hijack your attention, because their business is how long you stay. But what I actually want from them is to engage: talk to the right people, grow my audience, and get out.
When the AI that sorts your information works for you instead of the platform, it can rank your feeds by your goals instead of by engagement. It knows what you're working on and who you want to reach, so it surfaces the ten conversations worth your time and skips the rest. Or, like the contractor, it hands you the day on paper and lets you stay off the screen entirely.
I already do a version of this. My AI finds the posts worth replying to, and I spend twenty minutes replying instead of an hour scrolling. The AI finds and filters. I do the talking.
Security and reliability
As I share the intimate details of my life with AI, it gets uncomfortable that all of it goes to Anthropic. And it isn't just me. Running Dream Interpreter AI, I've seen how much personal and intimate detail people share with AI. People have written down over a million dreams there: their fears, their families, their grief, things many of them would never tell another person. Some of my friends now run their own local AI, with uncensored models, for exactly this reason.
Like every app today, Dream keeps those dreams on its own servers. The closest I can get to handing people control is an MCP connector, so they can bring Dream into whichever AI they already use. Letting people keep their dreams on their own server would be the next step, but for almost everyone, that server doesn't exist yet.
Privacy may not be what makes most people move, but security and reliability will keep getting harder to ignore. Attackers are handing more of the work to AI agents, and every big database of personal data is a bigger target than it was a year ago. Outages hurt too. On November 12, Microsoft's Exchange Online went down and people in dozens of countries lost Outlook and Teams for most of a day. It wasn't even an attack. If your mail lived on your own server at home, you'd still have every message you've ever received, readable and searchable, while everyone else waited for Microsoft.
It's already partly real
I've been living a small version of all of this. As a solo founder, I've spent far too much of my time in the details of building. Now I'm shifting my AI focus to much more than code: my marketing, my finances, my time, my energy and my relationships. That's a lot to keep track of, and as my sessions and memory files piled up, Claude Code kept losing track of what I'd told it. My rules were scattered across config files, memory folders and project docs. I called it harness rot.
So I moved all of it into my own planner app, on a server I control (in the cloud for now, not in my home yet), and connected it to my AI through MCP. It now holds my tasks and plans, my calendar and trips, my journal and notes, the people in my life, my body scans, and my AI's memory and rules. Every session reads my rules first, and every correction gets written back with the date and the reason.
Recently, Claude's phone and web apps couldn't connect to it, because they need a sign-in my server didn't have. My agent added one in a single session. Now Claude Code, Claude's apps and Grok all read and write the same memory, so there's no cost to switching between agents.
The same setup turned a photo of my InBody scan at the gym into a working body-composition tracker in an afternoon, with my AI as the only interface. None of this needed a startup's permission. It's my data, on my server, with whichever AI I choose.
The box at home
So here's where I think all of this goes. Everyone will have their own server at home. It holds all of your data, and it runs a local AI model that knows you. Your devices stop talking to dozens of companies' servers and talk to one server instead: yours. Your fitness band syncs to your phone, and your phone syncs to your box instead of a company's servers. Your phone can even run a small model of its own with a copy of the data it needs, so a lot of things never have to leave your pocket. Your bank, your messages and your feeds come in through it. If you hold crypto, your keys live there too.
It's a gate for your digital life, sitting on your home network. It takes in data, like messages, readings and transactions, but never runs someone else's code.
With a local model and a copy of Wikipedia on the box, you'd no longer need to access the wider internet for the vast history of human knowledge, only for what's new. And even for what's new, your agent can do the checking. Your box makes one call an hour to each social network, your email and the news, and keeps what's new and relevant to you, instead of your phone constantly preloading whatever will hijack your attention the next time you pick it up. Then it brings you what matters, maybe on the contractor's sheet of paper, so you don't have to be constantly connected. And yes, there's a zombie-apocalypse bonus: add a solar panel, and your box still holds a lot of what humanity knows.
Why now
A few years ago this would have been a fantasy. Four things changed.
First, the models. Today's frontier capabilities will be tomorrow's local capabilities. Epoch AI estimates that the best model you can run on a single consumer graphics card is only about 6 to 12 months behind the frontier, and the best open models overall trail it by about four months.
It doesn't have to be all or nothing, either. Think of it as three levels. A small model on your phone and laptop handles the quick stuff, instantly and offline. A bigger model on your box handles most of your life, because it knows all of it. And the frontier models from the AI labs only get called when a problem is hard enough to be worth the tokens. Because the box sits in the middle, it decides what leaves the house, sending only what that one problem needs.
Second, the economics. There's a cost problem hiding under today's AI subscriptions. Heavy users can cost an AI company far more than they pay: when Anthropic added weekly limits to Claude Code, it pointed to people running it around the clock, and one user ran up tens of thousands of dollars of usage on a $200-a-month plan. That kind of pricing can't last forever, and limits, prices and terms can change overnight. If your whole workflow depends on one company's servers, you can be rug-pulled. If you have local models, and you know what they can and can't do, you have a fallback that no one can take away.
Third, software is cheap to build now, which means you can fix what vendors won't. The official Gmail connector for Claude can't connect several Gmail accounts at once, and people have been asking for it for months. My agent could build that in an hour if the data lived with me. Vendors build one version for everyone. Your own box can build for you.
Fourth, AI can run the box for you. Home servers stayed a hobby because someone has to set them up, update them and fix them when they break. That's exactly the kind of work agents are now good at. My agent added a sign-in to my server and built the blog you're reading, each in a single session, without me reading the code. The tinkerer's job is becoming the agent's job.
And the plumbing to connect it all is finally showing up. MCP gives AI a standard way to plug into your data and tools. ActivityPub, the standard behind Mastodon, lets social feeds flow between services instead of being locked inside one app.
The Homebrew Computer Club days
This already exists, sort of. People run home servers today, and some of my friends run local models on them. But setting it up means picking hardware, installing software, and figuring out networking and backups yourself. It's for tinkerers. That's why I think we're in the Homebrew Computer Club days of personal AI: in the 1970s, hobbyists met to show off computers they'd built from kits. It took a consumer company, Apple, starting with a computer Steve Wozniak showed at those meetings, to turn that into something everyone had at home.
This has been tried before. Sandstorm, FreedomBox and Helm all tried to make personal servers easy, and none of them went mainstream. Umbrel and Start9 sell plug-and-play home servers today, and Umbrel can even run local models, but they mostly reach people who already self-host or run bitcoin nodes. What's different now is the AI. A home server used to mean hosting your own email. Now it means a model that knows your whole life and works only for you.
A real consumer version has to be the easy choice, something you'd bring home the way people brought home their first Apple computer: setup in minutes, automatic encrypted backups, getting your data back when the box dies, and an AI that works better because it knows your whole life.
Right now most people meet AI as something happening to them: a data center going up near their town, or something coming for their job. Local opposition blocked or delayed about $68 billion of US data center projects in just one quarter this year. I think people would feel very differently about AI if it were something physical in their home that made their life better and gave them more privacy online.
The hobbyist version of that future works today. Nobody has built the version for everyone yet: the Apple II of personal servers. A box you plug in at home that holds your whole digital life, runs your own AI, and makes every device and app work for you.
A note on how this was written: the ideas, stories and opinions are mine. I wrote this essay together with Claude, which helped me outline it, check facts, find sources and tighten the wording. That felt fitting for a post about using AI for more than code.