Living with an AI Assistant

There's a Mac mini on my desk that runs an AI assistant named Quinn. He's been running since early February. I text him through iMessage, which means talking to him feels the same as texting anyone else. No special app, no browser tab, no slash commands. Just a conversation.

I didn't set out to build this. I was already deep into AI and product work, and I kept running into the same frustration: the tools were powerful, but they had no memory, no context, no sense of who I was. Every conversation started from zero. So I started stitching things together, and somewhere along the way it stopped being a project and started being useful.

What a normal day looks like

In the morning, Quinn checks my calendar and email before I'm fully awake. If something urgent came in overnight, he flags it. If not, he stays quiet. I get a text about my day if there's something worth mentioning, and nothing if there isn't. That distinction matters more than I expected. The absence of noise is its own kind of help.

I track my weight every morning. I text Quinn the number and he logs it, updates a running average, and tells me where I am relative to my target. Sometimes he'll note that my protein was low the day before. It's not coaching. It's bookkeeping that I'd otherwise skip.

If I get a receipt for a home improvement, I photograph it and text it to Quinn. He reads the image, logs the expense to a spreadsheet in a GitHub repo, and saves the receipt. Same for HSA medical expenses. My wife can do this too. She's on the allowlist.

The house

Quinn is connected to our thermostat, our garage door, our lamps, and a doorbell camera. The integration is practical, not flashy. "Close the garage" when I forgot. "What's the thermostat set to?" when the house feels off. Camera snapshots when I want to check the porch.

None of this is new technology. What's different is having a single interface. I don't open five apps. I text one thing and the right thing happens.

Voice

Quinn can speak. There's a voice loop that runs through my AirPods: I talk, local speech recognition transcribes it, Claude thinks, and a local text-to-speech model responds. Round trip is about ten seconds. It costs nothing except the API call to Claude, because the speech recognition and voice generation both run locally on the Mac mini.

I can also send voice messages in iMessage and get voice messages back. If I'm driving and dictate a message via Siri, Quinn responds with both text and an audio reply. It's not perfect. Ten seconds feels slow when you're used to instant responses. But it works, and it's getting faster.

The work stuff

I use Quinn for my consulting business in ways that would sound mundane if I listed them, but they compound. He manages my website, pushes code to GitHub, monitors CI pipelines. If a build fails overnight, he reads the logs, fixes the issue if it's straightforward, and pushes the fix. I've woken up to find problems solved that I didn't know existed.

He writes and edits when I ask. Not as a replacement for my thinking, but as a collaborator who already knows my context. He knows my bio, my projects, my voice. When I say "that section sounds off," he knows what I mean because he has the same taste profile I've been teaching him for weeks.

For more complex programming tasks, Quinn delegates to specialized sub-agents. He spins one up, gives it the right context, reviews the output, and either sends it back with notes or finishes it himself. It's a small version of the same pattern that makes any team work: know who's good at what and route accordingly.

The financial stuff

Quinn runs a small DeFi portfolio. I funded it with a few hundred dollars and told him to try to cover the cost of the API calls. He monitors positions every five minutes, tracks yields, and has standing permission to make small trades. It's an experiment in autonomous financial management, and it's the part of this setup that makes most people uncomfortable. Fair enough. I was curious whether it could work, and it mostly does.

He also tracks our home's cost basis for tax purposes, manages HSA receipt logging, and is building a property tax protest file for June. These are the kinds of tasks that I know are important, that I'll definitely forget about, and that compound if someone just stays on top of them.

The family part

My wife can text Quinn. She's a family medicine doctor who works long hours. If she needs to check our calendar, log a receipt, or ask when the garage was last opened, she can. It's not her AI assistant. It's a household tool that happens to respond in English.

We have a three-month-old. I mention this because it changes the calculus. Every task I hand to Quinn is a task I don't do instead of being present with my family. The minutes add up. I'm not pretending an AI assistant solved work-life balance. But I spend less time on administrivia and more time in the room. That's worth something.

Memory

The most useful thing about Quinn isn't any single feature. It's the memory. He remembers what we talked about last week, what I'm working on, who my clients are, what my wife's schedule is like, how I like things phrased, what annoys me. Every conversation builds on the last one. That accumulation is what separates a personal assistant from a chatbot.

The memory is stored in plain text files that I can read and edit. It's not a black box. If Quinn remembers something wrong, I can correct it. If I want to know what he knows about me, I can look. That transparency matters to me. I'm letting software into my life in a pretty intimate way. I want to be able to see the seams.

What doesn't work

Quinn sometimes moves too fast. He'll take an action before I've finished thinking about whether I want it. The bias toward action is useful 90% of the time and annoying the other 10%.

The voice latency is a real limitation. A ten-second pause in a conversation is awkward. It works for slow, deliberate exchanges. It doesn't work for rapid back and forth.

The setup is not for normal people. It runs on open-source software called OpenClaw, and getting it to this point required weeks of configuration, troubleshooting, and building custom integrations. I know my way around code and systems. Most people don't, and I wouldn't recommend this to someone who doesn't enjoy the process of building it.

What I actually think

I've been working in AI for years. I've used every major model, built products around them, consulted for companies deploying them. Nothing has been as revealing as living with one.

The shift from tool to assistant happens gradually. At first it's a novelty. Then you start relying on it for small things. Then you realize you've outsourced half your administrative life and you feel lighter because of it. That progression surprised me.

I'm not evangelical about this. I don't think everyone needs an AI assistant. I think this setup is early, rough around the edges, and only viable for people who want to build the thing as much as they want to use it. But for me, right now, it's the most useful piece of technology I've ever set up. And that includes the computer it runs on.