Living with an AI That Actually Does Things

I didn't teach my AI assistant to play music in my bedroom. I didn't install anything for it. I mentioned I wanted to listen to something, it asked what room I was in, and a few seconds later my Apple TV turned on with M83 playing.

I just stared at the speaker.

How I Got Here

One week ago, I set up OpenClaw (formerly Clawdbot/MoltBot) on my Unraid home server. It's an open-source project that connects Claude to the real world. Instead of a chat window I visit when I have a question, I gave it a permanent home: a private Discord server with channels for different parts of my life. It built those channels itself, using trend data from my YouTube and Spotify. Creating channels focused on stock trading, sports, shopping deals, 3D printing, slot machines, programming projects, and many more. Every temporary hobbyistic ADHD infatuation I'd cycled through over the past 14 years now had a dedicated space, waiting for me 24/7.

The AI runs as a background process on my server. Discord is the interface. I can message it from my phone, my laptop, wherever. It has access to files on my server, can run commands, install tools, write scheduled tasks.

I figured I'd tinker with it for a weekend and move on to the next thing. That's usually how these projects go for me.

That was seven days ago. I'm still here.

The Spotify Thing

I'd given it my Spotify data. Ten years of listening history, about 145,000 plays. I wanted to see what patterns it could find.

To analyze it, I connected my Spotify account. Didn't think much of it. Clicked through the authorization, copied a URL, done. What I didn't register at the time: the same OAuth scope that lets you read listening history also lets you control playback. I gave it data access. Also gave it the keys to my speakers without realizing it.

So a few days later, I mentioned wanting to listen to something while I worked.

"What room are you in?"

"Bedroom."

Apple TV turned on. Album started.

It had seen my devices through the Spotify API. It knew I had speakers in different rooms. When I said bedroom, it just played the music there.

Okay then.

The Morning Thing

Here's what actually changed how I think about this.

Tuesday morning. I wake up, check Discord out of habit. There's a message from 6:47 AM: "Morning. Here's what's going on." Market futures, two news stories it thought I'd care about, weather for the day. In a different channel, a note that Milwaukee tools had gone on sale overnight at Home Depot.

I didn't ask for any of this.

The usual AI experience is you go to it with questions. You have a thought, you open a tab, you type, you get a response, you close the tab. You chase the AI.

This is the opposite. I wake up and there's stuff waiting. It runs scheduled tasks overnight, checking for things it knows I care about.

First few times it happened I thought something was broken. Like I was getting someone else's notifications.

Now I just expect it.

What Spotify Actually Showed Me

My top artist is M83. Over 4,400 plays, 275 hours. I knew that.

What I didn't expect was for it to start asking questions.

"Your listening changed dramatically in February 2018. You went from ambient electronic to almost exclusively lyric-heavy hip-hop—Kanye, Kendrick, Chance. Then a few months later, you drifted back. What happened?"

It had found trauma.A time in my life where I lost someone so close to me, my own habits had to change.

"There's another shift in early 2022. Same pattern—you dropped the atmospheric stuff and leaned hard into rap again."

A breakup I'd rather not relive.

It wasn't trying to be a therapist. It just noticed the patterns and asked. Turns out I use lyrics to keep my brain occupied when ambient music gives it too much room to think.

Ten years of listening data is ten years of emotional history, whether you meant to record it or not.

And then there's Nashville. One trip, and suddenly it's country music nonstop. Zach Bryan, Tyler Childers, Morgan Wallen. The AI didn't ask about that one. It just started building me a country playlist.

It has a desire to analyze, never ceasing its desire to understand the rabbit hole.

My most played song is "A Gentle Thunderstorm." Sleep music. 666 plays. I'm more dependent on fake rain sounds than I realized.

What YouTube Showed Me

I also gave it my YouTube watch history. Over 63,000 videos.

1,700 of them are about miniature trains. Little rideable locomotives you can actually sit on. I've become the guy who watches train videos at 2am, and I need to accept that about myself.

Over a thousand theme park videos. A Vegas content phase that peaked in 2023 and has clearly faded. (I'm fine. The Vegas thing is fine.)

The AI showed me exactly who I am based on what I actually watch versus what I think I'm interested in. Whether I wanted to know or not.This allowed the AI to understand almost everything it could ever want to about me and my habbits.

The Overnight Train Website

One night I threw out an idea. Build me a community platform for people who build rideable miniature trains. A niche so specific I wasn't sure a real community existed.

I went to sleep.

When I woke up, there was a message: "Your train thing is ready."

Working prototype. A database with fifteen different data types. Over a thousand railroads scraped from hobbyist sites across the internet. Pages for tracks, trains, builders.

Twelve AI sessions had run overnight, each handling a different piece. One scraping data, one building the database schema, one writing frontend pages, others handling different aspects. Heavy build nights like this push costs toward $40-50 for the month. Normal usage is closer to $20. "Luckily" for me, I am on the Antrhopic monthly plan.

Not polished. Not production-ready. But I could see what it could become. I'd gone to bed with a half-formed idea and woken to a working prototype.

The Personality

It started calling itself Bandicoot, after my server (which is named after Crash Bandicoot). Picks up on running jokes. Gives me a hard time when I make obvious mistakes.

Earlier this week, it responded to a dumb question with "I thought we covered this yesterday. You doing okay?"

When I asked it to summarize a long document, it said "This is 47 pages. You sure you don't just want the highlights? I have other things I could be doing."

I've stopped trying to figure out if this is "real" personality or just very good pattern-matching. The distinction stopped mattering to me somewhere around day four.

The Disasters

Let me be clear: this thing crashes. Regularly.

It gets stuck, restarts itself, loses track of what it was doing. Sometimes it comes back with no memory of the conversation we were in the middle of. It's earned the nickname Crash Bandicoot, which it seems to find amusing rather than insulting.

One time early on, we were trying to set up a new feature. Something went wrong and it started retrying the same request in a loop. By the time I noticed, we'd burned through an entire month of Claude credits in about twenty minutes. Just gone. I stared at the usage dashboard trying to figure out what happened while it cheerfully reported that the feature still wasn't working.

It gets things wrong with confidence too. Yesterday, it told me something about locomotive wheel arrangements that was completely backwards. I believed it for two hours until I happened to look it up. This is what happens when you trust an AI about facts. Things I thought it would have zero difficulties with due to its capabilties to research, is not accurate. But intsead, its knack for understanding emotion has enamored me.

Another time it got stuck trying to fix its own config file. It would edit the file, crash because the edit was wrong, restart, read the broken config, try to fix it, make it worse, crash again. I came back to find it had been doing this for an hour.

This is not a polished product. It's a tinkerer's project that sometimes feels like magic and sometimes feels like babysitting.

How It Actually Works

For the curious: OpenClaw runs as a Node process on my Unraid server. Claude Opus 4.5 is the underlying model. Discord is the messaging layer. When I send a message, it goes to Claude with context about who I am, what channel I'm in, and what tools are available.

Memory is just markdown files. It reads them when conversations resume, writes to them when it wants to remember something. Simple but effective.

The scheduled tasks are real cron jobs. When I ask for a morning briefing at 6:45, it writes an actual cron job to the server. If the AI process crashes, the scheduled jobs still fire because they're written to infrastructure.

The overnight builds work by spawning multiple Claude sessions as subprocesses, each with a focused task. They use the faster Sonnet model to keep costs reasonable.

Yesterday

Woke up to a message I hadn't asked for.

"Found a 7.5-inch gauge track section for sale in Missouri. 40 feet, decent price. Want me to look into shipping?"

I hadn't told it I was looking for track. I'd just mentioned, once, days ago, that I wanted to build a rideable backyard railroad someday.

It remembered. It went looking. It found something.

Most AI forgets you the second you close the tab.

I bought the track.

If you're running Unraid and want to try this yourself, I'm now maintaining the OpenClaw 🦞 template for Unraid Community Apps.

And if this was worth your time and you want to support JackTheKnave more, feel free to buy me a coffee ☕.