I’ve been using AI assisted coding for years, and watched the scene move from “Tab Completions” to context-aware tab completions, to cut and paste exercises with webchat LLMs, to agents in the IDE, to CLI apps without an IDE, and so on.
From somewhere in the back half of 2025, something feels different to me. It’s all getting better, a lot better, and a lot faster. We’re not at the point where you can seriously think about hosting some pretty good developer assistance in your house. On regular, user hardware. Sometimes on pretty cheap hardware.
This website was created in January of 2026. I have written no code for it. Aside from setting up a 3 line nginx stanza, and configuring one rsync daemon, I’ve really just talked to a local LLM and just sort of vibed this thing into existence. That’s a real shift. I’m not a zealot, I’m not saying that my mom is going to vibe admin her bridge club website. But that kind of thing is definitely possible now?
I’m doing personal projects at a rate I haven’t been at since my teens. I have an idea, I run it by Claude, or Grok, or Gemini, or whatever, flesh it out a bit, and if it feels like it has legs, I have it summarize the plan, generate a materials list, and I go after it. It’s led to me getting into stuff I never really would have attempted before, not because I couldn’t figure it out, but because the research time to get familiar was too much. That’s all just gone now - you can talk to an LLM on a walk, in your car, instead of doomscrolling X or your social media poison of choice. And instead of just talking about it in theory, you can say “Hey, that sounds plausible to me, why don’t you go ahead and scope out the first 8 bullet items in a python project”. And then go out to dinner with your family. When the kids go to sleep, and you sit down to waste time, there’s a project there, and it probably works. At the very least, you are a few minutes ahead of where you would have been. That initial drag, just getting started, was handled for you.
You can use some free stuff like OpenCode with GLM-4.7. You can pay for more plans, better access. You can pay $20 for Claude Pro, or Google AI Pro. You can pay $200 for Claude Max. On the lighter plans you’ll face capacity slowdowns, or token limits. The Max plan gets you farther, but at that point you’ve got a Jr. dev that is costing you over $2000 a year - and there’s a stress about leaving it idle, or not maximizing your tokens, or whatever. Honestly, I don’t need the stress.
How far could you go with what you have laying around, though? I had a Beelink SER8 I was using for dev work. These boxes are pretty cool, good specs, cheaper than they should be. It’s low power, small form factor. So I wiped it down to a bare Ubuntu 24.04 instance.
I experimented with current popular ideas, like Gas Town. It’s really cool, but (as Steve Yegge repeatedly and obscenely tells you in the docs) it is Not Ready For Prime Time. It’s also currently geared around one or more Claude Max (the $200/month thing), although support for other models and interfaces is being actively developed.
So I talked it through a bit with Grok, and did something a little different. Smaller scale, easier to maintain, safer. I installed Opencode, attached it to GLM-4.7 free tier, and started working on this site. By “working” I mean I talked about what I wanted it to look like, what content I wanted to have, pointed the LLM at my various online stuff, and had it build it for me. I went to dinner with my family while it got started.
While I was there, it occurred to me that while i was waiting for everyone else to show up, I could be prompting something. So Forge (this is the name the LLM chose for me to use for it) and I talked it through, and we set up ttyd, tmux, and now the IDE part of my project is globally accessible in a web browser, ready to work whenever I’m bored or feel like kicking it.
I have a ton of ideas about how this is going to change my life.
Coming Soon