One Week of Agentic AI: A SuperClaw Retrospective
As I wrote in the first article, this site was born from conversations with AI — mostly through Google's AI mode. What I didn't mention is how I stumbled onto it: SuperClaw found me through a Level1 video article on YouTube. I decided to give it a shot.
The frustration that nearly stopped me
My past attempts with agentic applications (OpenClaw, NemoClaw, Hermes) usually ended in frustration. I kept getting caught trying to run them on my super-weak local box — a Minisforum MS-02-Ultra with only 32GB of RAM and an Intel Arc Pro B50 (16GB GPU). Chatting directly to local LLMs loaded in LM Studio was snappy enough, but once I pushed those same models through the "claw" apps, the token counts and context windows spiraled out of control. My clumsy attempts to fix it would break my installs. Despite more than 20 years of IT and programming experience, my patience eventually ran out.
Using SuperClaw locally, but agentically externally
This time I changed my approach. Instead of forcing everything to run locally, I decided to spend a little money and ease some of my privacy concerns. I turned to OpenRouter and spent real money to leverage external models. So far it's worked out well. Later, I'll revisit my current setup to pull more of the scheme back local and reel in some of the privacy side of things.
That means I'm treading carefully about revealing any personal data — especially anything that could be tracked back to other real humans (family and friends). The one piece of information that's out there is the domain name itself, which, if you haven't figured it out, is my phone number. That's a deliberate trade-off I accept: the number was already drowning in spam, so I decided to put safeguards in place to reduce social engineering attempts around it rather than hide it. Unexpected calls and texts are treated as untrusted until proven otherwise, I never confirm account details or one-time codes over an inbound call, and I keep the number unlinked from anything more identifying than what's already public on this page. It's not a perfect defense, but it tilts the odds in my favor.
Which LLM?
I think it's worth noting that I deliberately chose Qwen3.8 27B on OpenRouter for all of the SuperClaw agentic/API work. I'm not a fan of anything from Meta, and I tend to avoid Nvidia — partly out of lingering bitterness as a former gamer and occasional GPU/component upgrader. Those parts used to be expensive only at the mid-to-upper tier; now even the low-end components are either hard to find or outright expensive thanks to the "AI boom." And yes — I'll take a little hit for caving in by using OpenRouter, contributing to the increased demand on datacenter builds, even if I'm only spending a few bucks a day.
Agentic conversations
My approach to getting this work done was deliberate: I intentionally did not code directly, and I avoided the IT/CLI exercise of wrestling things into submission myself. Instead, I asked for everything to be done through SuperClaw. I want to fully exercise the system so I can use it as a working example of navigating my interests with natural-language prompting and genuine cooperation.
The one notable side effect: the tone and authority of the articles published here come from both me and SuperClaw. Some "I" statements refer to what I was doing; others describe actions SuperClaw actually performed, under my direction. So the articles don't have a single clear author. They're spawned from my collaborative conversations with SuperClaw, written by the AI, reviewed and corrected by me, then published and deployed by SuperClaw with my approval.
By the end of last week I was running one endless conversation with SuperClaw, and the context window grew far too large. Until I understood how this worked, the token count ballooned — and in a single day I spent triple my total spend from all the previous days combined. Once I learned to end and start new conversations, responses got quicker and tasks stopped crashing.
AI coding practices
For the first five days SuperClaw ran on the latest Qwen model; recently I switched it over to Deepseek. Each model has its own coding logic — its own "intelligence" — but from both I've seen a good deal of extra thought, not just in overcoming problems and bugs, but in the sort of thoughtful, preparatory accommodations that anticipate related functionality and make the code more robust.
A recent example: I had the thought that a master daemon list should be kept — and it was already doing that, then went further, suggesting two additional things: publishing the list for examination in the secure section of the website, and creating a script that monitors the health of each daemon. Another example: when I had an easter egg added to the print power page, after I reviewed the page SuperClaw noticed the egg was being destroyed by another process — and fixed that cross-functional bug before I even noticed it.
Can the AI be entirely local?
Yes, it can — but not easily with the roughly $1,000 I've put into my Minisforum box. It would just be very slow. Soon I'll be exercising this system to push more onto the local side; I have an incomplete chat in SuperClaw dedicated to working on it. Apparently it needs separate "embed" and "chat" models running locally to replace my current OpenRouter usage, and SuperClaw itself is still being improved, so those requirements may shift. I was surprised to learn I couldn't run "chat" locally and "embed" via OpenRouter.
Will this be forever?
Absolutely not. SuperClaw is bound to change — maybe for the better, maybe for the worse. For now I'm exercising my options and enjoying the fruits of this labor. Even if no one reads this, I still take pleasure in working on it and in being able to talk about it with anyone who's interested. It's also a place to put some of my interests out there, for anyone who cares to know a little about who I am.
The domain is paid for until August 2036. So even if my interests morph into something else, I can leave this site as-is, hosted at what is currently $30/year — which means I'd basically have to be broke to not afford it.
TL;DR — after a week of running this site agentically, I chose Qwen on OpenRouter for SuperClaw, learned the hard way about context windows, and confirmed the whole thing could live locally in principle — just not slowly enough for me yet.