So - I’m planning on spending Easter on trying to let an AI agent live “close to home” - with access to files and certain freedom to operate on it’s own.
There’s a hype right now with Openclaw, but I am looking at others as well - lately the focus has been Zeroclaw.
My main concern is information and system safety. I don’t want this thing to run wild. There is no “Ctrl+Z” / Undo if this goes bad.
I don’t want this thing too close on everything until I understand more - so I’m moving ahead with precaution (separate network, dedicated firewall - and separate hardware).
(Yes, I am aware I could just host externally for reasonable price to achieve isolation - and that local hardware does not provide safety if misconfigured - but going in the cloud does carry with it some tradeoffs that I do not like).
I do see some connections and applications between personal agentic AI and JD.
Example: I have some “boring” Areas (Routines Maintenance Logs) that don’t get the attention or structure they deserve (as I am probably the wrong type of person to provide them the proper attention). I do see that a meticulous sub-agent could be assigned to keeping up with this, if I only feed in the data. I am the type of person who appreciates the result of such work.
Has anyone else been playing around? Any first impressions on combining JD with agents/sub-agents?
PS! I am not advising or recommending anything at this stage. I am still in early learning stages, and it’s based on personal curiocity.
I don’t use a local ‘claw’, partly because I don’t have a need for that, and partly because it feels very early and a bit dangerous. But obviously it’s where things are going.
For me, Claude Code at the Terminal does everything I need it to do. I prefer prompting and watching vs. having anything run unattended.
But I do hook Claude up to my vault. We have a video scheduled for Wednesday. Keep an eye on the channel.
I used open claw. It was amazing and helped me get rid of the backlog. Understand the concerns you all talk of but the amount of time it saved me was wonderful.
HP MiniPC as “Openclaw host” (32 GB RAM, i5-9500 CPU, no GPU) running Debian 13 Xfce
Using Logseq for documentation of “steps” and learning as I go along (Logseq graph will also live inside the Openclaw host, so we (me and Openclaw) can share the information)
Status as of “Palm Sunday” 2026:
Firewall done
Host done
Next is install of Openclaw itself
(Iterations on letting Openclaw play with JD principles comes a bit later…)
Tip:
Recently discovered Fabric (patterns) Worth checking out if you are getting into tinkering with the Agentic AI space.
No reason to re-invent the wheel (at least not every time).
EDIT** Claude is much better. I downloaded the GitHub repositories and along with the manual used Claude to write terminal scripts to help categorize and move files. Much better than open claw (and cheaper unless you use your own repository which I couldn’t be bothered to do).
Claude did advise against some of my deviations from the categorization in the new jdex rules (sub categories). It works but I wonder if it’s a massive issue having two decimal points? Ie 11.34.1
Main target of “Easter session 2026” was to try to “catch up” on AI knowledge - getting my bearings, as I felt the train was leaving the station without me.
Didn’t take very long to realize that a CPU only hardware is not capable of handling even small, local LLMs.
Did some exploration with various cloud providers and different subscriptions.
That evolved into one of the biggest rabbit holes I’ve ever fallen into - but I guess that is everyone’s “getting into AI” experience at the moment.
This again led to the procurement of a fairly heavy consumer hardware that is capable running local LLMs. Inputs to this strange decision (which caught me by surprise) were mainly; chance of short term hardware price drops vs spending token/subscription money vs opportunity cost of waiting vs keeping control of data while I still don’t know what the heck I’m doing.
I am now +20 days after the hardware arrived, and this thing has legs/potential.
Target now is to achieve sufficient knowledge to be able to separate toy from tool, hype from actually useful. And where that might lead next, I seriously don’t know. I am not naive, but curious.
JD integration and automation is definitely still on the list, but the list has become long with a lot of competing priorities.
