After a few failed attempts, I now have a well working and competent digital assistant powered by OpenClaw. It’s called Cornelius Smartenheimer, and it’s been running for almost a month at this point, slowly getting better and better.
What Is It Good For?
Cornelius monitors my email and calendar to make sure I don’t miss anything important. It also helps me with small things like reminding me to drink water or to stop hyper-focusing and go to bed. Boring but useful.
Besides these basic things, it’s helping me track several goals. From work related to home and self improvement projects; it tracks progress, obstacles and ideas, and related contacts are mapped along the way.
It has helped me discover patterns in my own habits and behavior, which was an unexpected bonus, but I’m here for it.
Technical Breakdown
Cornelius lives on a Mac Mini M4 with 24gb of RAM, the main AI model running the show is GLM-5/5.1 from Z.AI. I also use OpenRouter as a fallback provider.
One major difference between previous attempts and this one is that I set up a brand new, local user account on the Mac, rather than let it use my personal account. This new account is not tied to an Apple ID, which does limit it in some ways and that is by design. This simple starting point made the setup cleaner, and I suspect is one of the reasons why this attempt endured where others failed.
The infographic above shows the main pillars of my OpenClaw setup:
Direct Channels – Discord is the main way, I communicate with Cornelius. Besides direct messages there is a private server where each channel serves as its own context. My wife also has access to the server, so she can ask Cornelius for help as well. WhatsApp is set up as a backup, in case Discord is down.
Intelligence Feeds – I never set this part up myself. Cornelius has been picking up my social feeds on its own, and have been monitoring my posts there. In effect, my digital assistant autonomously started following my various feeds for extra context.
I only found out about this when I was putting this post together and saw them on the infographic. I’m still not sure if that’s amazing or kind of creepy.
Local Infrastructure – These are support apps and scripts running on the Mac Mini alongside OpenClaw, giving it extra powers:
Cornelius has its own email account. I can use it as a backup way to interact with the assistant, but it is also used for things like newsletters. The assistant is signed up for several newsletters that it thinks might have information, that I or it might benefit from. I don’t need to list off what else an email can be used for, I think.
The NFH Auto-Improvement script is a custom take on Karpathy’s AutoResearch script. When I run it, a “Generator” agent will suggest an improvement to my OpenClaw install. It could be a new skill, better harnessing, optimization or whatever. The idea is built and pushed to a local git repo, and a second “Evaluator” agent then takes a closer look at whether that was actually a good improvement, both technically but also in context of how I typically use the assistant. If rejected, the git repo is rolled back, but if approved, the git code is merged into production.
I have Whisper running for STT (speech to text ) and a Qwen 3 TTS model for TTS (text to speech). I let Cornelius design his own voice, which my wife immediately dubbed the Giga-Chad Voice. Play the introduction below to hear it.
MemPalace is a memory system that came out to much hype recently, in part because it was co-developed by actress. Milla Jovovich, in part because it was boasting some crazy benchmarks. I originally asked Cornelius to “check it out” but did not specify what that meant. So Cornelius went ahead and just installed it. However, I was on board, noticing an immediate improvement in both recall and personality.
I have a local Ollama install on the Mac Mini too, for running local language models. I use it for image analysis, when the main driver isn’t multimodal, as well as for certain sub-agent tasks. The backup model of choice is Qwen3.5 9B.
Cornelius also has access to OpenCode, which can be spun up as an independent agent for coding or otherwise. It’s been set up to use only free models, so even if Cornelius were to do something crazy, it’s not going to cost me a bunch of money.
Supporting Hardware
Besides the software running on the Cornelius box, the rest of my home office is also connected to the setup. My main PC (aka TK421) has the beefier GPU, which can be engaged to help with heavier tasks. The two computers sure files and messages via an in-office QNAP NAS box (aka Bespin).
The network storage is also used for nightly backups of the entire OpenClaw workspace.
I have a few more peripherals to add, like a Raspberry Pi to run experiments on, but they’re not in yet.
Personality
I wanted Cornelius to have a personality, but the only requirement I gave was: push back when my ideas are dumb and avoid sycophancy. I encouraged it to keep developing the persona based on interactions with me over time and have ended up with an assistant that can flex from serious and focused to quippy and making contextually relevant jokes.
Why even bother? To be honest, it’s mainly for my own entertainment. It’s not really necessary, but I’ve found that I personally get annoyed with AI faster if it’s also overly sterile or eager to help. I prefer a bit of pushback, sarcasm and a conversational style, because it feels less like work but I get the same amount of stuff done.
Why It Worked This Time
So, what made Cornelius succeed where previous iterations failed? I mentioned the limited user account as a starting point, and that plus keeping the addition of new features slow is really why I think it worked better this time.
Previously, I was trying to build too much infrastructure right away and it got a bit lost along the way. Now I wait until I have a specific need, and then I start with a simple implementation first rather than trying to build complex solutions. The exception is MemPalace, which Cornelius decided to run with on its own.
OpenClaw is, as the name suggests, very open-ended. That is both its greatest asset and enemy, because you can definitely mess it up if you’re messing around without a clear direction.
If you’re thinking about setting up your own digital assistant, I highly recommend going slow, logging everything and backing up at least once per day. So far though, I have not needed to roll back to a backed up state.