AI Psychosis is a thing, and I’ve got a mild case of it.
The shape of it is familiar to anyone who’s caught themselves doom-scrolling. Instead of doing the thing I actually need to do, I’m prompting. I’m vibe coding “just a little longer.” I’m having my agent reorganize my notes when I should be folding laundry. The dopamine of the prompt-and-response loop is real, and it’ll happily eat a Saturday if I let it.
At the same time, those of us in the tech industry are recognizing that we need to lean into AI to have relevant skills post-AI boom. There are some brutal economic forces at play, but making the best of it, I’ve taken that latent dread and turned it into some educational fun. And that is why I spent the last month renovating my personal AI setup. The goal of the renovation was to get from “I should be using AI more” to “AI should be doing more for me, quietly, while I do something else.” Less time staring at a prompt window, more time in the rest of my life.
From a propped-open laptop to the cloud
A month ago, Hermes lived on a spare MacBook called claw, propped open with Amphetamine, keeping it awake. The core model was Claude Haiku via the Anthropic API. That setup worked, but it had two problems.
First, my home internet flapping took Hermes down with it. Second, every interaction costs tokens. Only about a dollar a day on average, but the metering was psychological friction. I’d find myself thinking, “Do I really need to ask this?” and then not asking. The thing I’d built specifically to be useful was sitting underused because I was nickel-and-diming myself.
So I moved Hermes to a cloud VPS. Uptime is no longer my problem. While I was in there, I also reskinned the personality. Hermes is now Rex: a refined Tyrannosaurus butler. Same agent infrastructure, different character. Hermes was fine, but as the project has become more popular and overtaken OpenClaw, I started seeing her represented in more places, and I wanted something unique for my setup.
The split-brain
Moving Rex to the cloud created a new problem. A bunch of the things Hermes used to do on claw required either my home network or my home IP address. yt-dlp in particular gets rate-limited or blocked when run from a datacenter IP, and a few of my authenticated services aren’t exposed to the internet. Rex on a cloud VPS can’t do those things directly.
The fix was a split-brain setup. Rex lives in the cloud and is the general-purpose brain. I also run a locally-hosted N8N instance at home that holds the credentials and the IP-bound tools. Rex can call N8N workflows but can’t modify them.
That last constraint is the one I’m happiest with. It means I can let Rex go off and vibe-code solutions to problems without worrying that it’s going to wire itself a backdoor into my authenticated services. The dangerous stuff sits on the other side of an interface Rex doesn’t own. If Rex wants new capabilities involving credentials or my home IP, I have to build those into N8N myself. It’s a clean seam between “Rex can play here” and “Dave has to think about this.”
Swapping where the tokens come from
While I was at it, I changed where the money goes. At work, I’m still using Claude Code - it’s not going anywhere, it’s too good - but for personal use, I turned my OpenAI subscription back on. Rex now leans on OpenAI’s Codex quota for most of its day-to-day work, which means I’m paying a flat subscription rather than per-token API billing.
The practical result: I’ve stopped flinching when I ask Rex to do something. The metered tap is closed. That alone has changed my behavior more than the personality change or the cloud move.
Subscription rates for LLM access is likely a heavy loss-leader for the big AI companies. The street value of the GPT 5.5 tokens I’m burning is about $30/day, which I’m effectively paying $3 for. A coworker of mine said recently, “If you could go back to the days of $4 Uber rides, would you?” So yeah, ideally I’d like to not get trapped in an ecosystem … but I am very willing to take $4 Uber rides.
What Rex is actually doing for me
With the flinch gone and uptime fixed, the list of things Rex handles in the background has grown:
None of these is heroic on its own. Together, they’re the difference between a tool I have to remember to use and an assistant that’s already done the thing.
The big difference between Rex doing this and using Gems, Projects, GPTs, or some other construct of a big AI company is that Rex can create and revise its own skill instructions for each task and write its own code to automate processes or transition deterministic parts of what it is doing to non-LLM functions.
The fun one: Rex in the cockpit
The side project I’m proudest of right now is a voice agent kneeboard for Star Citizen. I have a push-to-talk macro on my Windows gaming machine bound to a web-delivered voice interface to Rex. Mid-flight, I can hit the macro and tell Rex what’s happening, and he keeps a flight log of the session. Cargo runs, combat encounters, who flew with me, what went wrong, what we should do differently next time.
It’s working better than I expected. Local voice models have gotten good enough that I can talk to Rex casually while my hands are on flight controls, and the log entries he produces are actually useful when I’m trying to remember what happened in a long play session a week later.
The other side of the psychosis
Is the renovation working? Mostly. The metering anxiety is gone, Rex is doing real background work, and I’m not juggling AI tools as much. But I’d be lying if I said I’ve solved the prompting-instead-of-living problem. I expect I will still catch myself building a new Rex skill when I should be outside.
The point of a personal AI operating system, I think, is to make AI feel less like a slot machine and more like plumbing – out of sight and mind I’m not all the way there yet. But having Rex earn his keep in the background, while I do other things, feels like the right direction.