To offer a counter perspective, I ended up somewhere completely different and it’s working for me.
First, to establish my credibility: I’ve got a Mac mini m1, a m2 MacBook Air, iPad, iPhone, etc. Buuuut I’ve also been running a low-power Home Theater PC/NAS for years that keeps all my movies, music, a PDF archive, Time Machine backups, photos, blah blah blah. It started as an experiment with Raspberry Pi’s and old hard drives, then graduated to some Intel NUCs I found, and for the last couple years a HP Mini Tower that sits next to my washer and dryer and is always on.
That machine is modest - not the fastest processor by any stretch, 16GB of ram. But it came with a 512 SSD and I put two 13TB drives in it - one backs up the other. It’s all run on Open Media Vault. Here’s the thing though, minus those two hard drives the whole thing cost $120.
And here’s where I get into the next level nerdiness – apologies in advance.
Most of the work in my workflows is done by Claude - not the local machines processor. As soon as I saw how useful this Robot Assistant thing was I, like you, wanted access and for it to be able to run any time. So I looked into NanoClaw and realized quickly: huh, this security stuff seems totally manageable. I took the leap and
- installed NanoClaw
- connected it to WhatsApp
- installed Obsidian
- hooked up Obisdian Sync
- set up some other MCPs for email, calendar, etc (with the help of Claude Code)
- told NanoClaw that the Obsidian Skills folder is the “source of truth”. A few skills are specific to that machine - like what to do with what’s coming in from WhatsApp – but it’s a handful maybe.
Around security, there’s a few details about isolating things with a NanoClaw user and containers, but Claude Code helped me get through all that.
So far it’s been great. I get a morning, mid-day, and evening updates sent to me via WhatsApp. I can send a message from my phone and say things like “I have 20 minutes right now and I have my phone, what should I do” or “I’m near the post office and have 15 minutes” and it gives me a good answer. I can be in my car, think of something, say it to WhatsApp via CarPlay, and it gets handled.
I even had Claude Code improve the server itself. Turns out I had TBs worth of dupes I didn’t realize it caught, fixed, and cleaned up.
Someday, when I can run the whole LLM locally, I wouldn’t mind paying for a beefier machine. But for now, when the AI processing is happening elsewhere, and I’m mostly working with calendars, email, PDFs, markdown, and other things that aren’t macOS specific or churning through video rendering projects, I don’t see the need because this is awesome.
Not for everyone, but I think worth considering.