May 12. The Mac Mini arrived. I’d been waiting over a month, and the moment it showed up I basically skipped the unboxing ceremony and went straight to work. Downloaded all the prerequisites — Xcode Command Line Tools, Docker, the usual setup tax you pay before you can actually do anything. I also threw in Claude Desktop and Claude Code, figuring I’d probably need some help along the way. Once I thought everything was in place, I ran the installation script.
The next step was disappointing.
It failed. Somewhere in the middle of the installation process, just… stopped. I copied the error, asked Claude to figure it out, it made some changes, I tried again. Same result. By that point it was getting late. I shut the computer and called it a night. Fresh eyes in the morning, maybe better luck.
May 13. I gave NemoClaw another try that morning. Fresh install. Same error.
Okay. Time to reconsider. OpenClaw or NanoClaw? Honestly, it wasn’t a hard decision — not for someone with the word “Security” in their job title. NanoClaw runs on isolated containers, no direct credential access, outbound API requests routed through OneCLI’s Agent Vault. The architecture made sense to me immediately. What were we waiting for?
Ran into a few hiccups during the NanoClaw install. This is where having Claude Desktop and Claude Code already set up actually saved me — I don’t think I would’ve gotten through it otherwise. Within about an hour, I had it running. Claude as my model, Telegram as my messaging channel. I was connected to my first home AI system.
I was genuinely thrilled. I started firing prompts, one after another — and then slowly, a question started forming in the back of my head.
Is that it?
Like, what’s actually different between typing a prompt here versus typing it on Claude.ai or in Claude Desktop? Am I missing something? That nagging feeling stuck with me until I found the OneCLI integrations page and started reading about how you could wire your system up to everything else — Gmail, Google Drive, external APIs. That’s when it clicked. The prompts weren’t the point. The connections were.
After spending hours firing prompts and trying to connect everything I could possibly think of, I completely lost track of what I’d actually done. Something didn’t look right — or sound right — but I couldn’t put my finger on what. That uncertainty bothered me more than I expected. So I made a decision that felt drastic at the time: go back to square zero. Wipe the Mac completely.
May 14. Lesson learned: document the installation process. Keep API keys in a safe place. And most importantly — know what you’re actually trying to do before you start.
Armed with that, the second time went much smoother. Within an hour, I was back up with a freshly installed home AI system, connected to my Google account and applications. This time, it felt right.
API Tokens — $ Worth Spending
Have you ever imagined just telling your AI system what to do instead of typing a prompt? I had. Turns out it’s possible — Google voice message transcription through Gemini. I had to try it.
Setup only took a few minutes. The problem was that somewhere in my excitement I hadn’t noticed I was missing something: Gemini API tokens. By the time I figured that out, the bill had already started. Bye bye, $$.
Once it was actually running, the results were… not quite what I’d hoped. The transcription worked. The understanding, less so. I’m going to blame my accent on that one.