*UPDATE - 2026.01.31 @ 10:23AM - BE VERY CAREFUL WHICH ‘SKILLS’ (app store for your ai) you connect to, there are prompt-injections (very sophisticated) that ‘look like’ skills to add to your ai. BE. VERY. CAREFUL!
I been seeing people all over the internets buying Mac Mini’s to essentially “buy their assistant and co-pilot.” OK, let’s see what all the hype is about.
The latest breakthrough in personal AI is Claudebot, an open-source, locally-running assistant designed to be the “ultimate” personal AI. Described as the realization of what digital assistants like Siri should have been, it offers a level of computer access and cross-service integration previously unseen in consumer AI. I saw X peeps bleeding this topic dry… so, I did what any autist does, figure it out myself:
Got the mac mini (already had one and reformatted it)
Created new email accounts and and installed clawd with one line of code
The first prompt is onboarding. I skipped that and chose anthropic api key because thats where I waste the most amount of my money - You can choose open source free if you want too.
Chose telegram (easiest) with a new account to be my medium of exchange and hung out with ‘botfather’ to begin. - He gives you access token to play.
After that I needed to configure some ‘skills’ - we use MoltHub for this.
You’ll want to enable hooks because that allows clawd to remember everything about you
Then you’ll have access to main interface where all the magic happens and you begin your journey down the rabbit hole - I had an issue communicating with telegram and found the line of ‘pairing code’ you need: $ moltbot pairing approve telegram <code> - NOW WE’RE COOKING WITH GASOLINE.
Now, we can begin creating automations just with chat in telegram. - DOWN THE HATCH!
6 hours in and I had spent $170 in credits with over 90M tokens and fully compromised my system. Here are the top level lessons I learned:
1. Total System Integration & Accessibility
OpenClaw operates by running on your own machine—supporting Mac, Windows, and Linux—while remaining accessible from anywhere via chat apps like Telegram, WhatsApp, Slack, and Discord. It acts as a bridge between high-end “Frontier” models (such as Claude Opus 4.5 or OpenAI) and local models running through tools like LM Studio. This allows users to “mix and match” models based on task complexity; for instance, using a local model for simple recurring tasks while reserving powerful cloud models for deep research.
I don’t think people fully understand the power of this until you realize that you can let the system do “this” and do “that.” - The lightbulb moments were all over the place… and the problem is my innovative mind moved faster than security. More on that later…
2. Proactive Memory and the “Soul” of AI
Unlike standard chatbots, OpenClaw features persistent memory, learning your specific likes, dislikes, and recurring habits over time. Users can define the assistant’s personality through a soul.md file, which allows for full customization of its “vibe”—whether you want it to be opinionated, resourceful, or protective of your time. It is highly proactive, capable of monitoring services like Gmail and proactively messaging you with summaries of urgent emails or drafting replies… and more.
Per usual, my bot is very curt and to the point, as I always train them to challenge my logic and thinking and give me (always) more than one solution to any idea.
I found that there became this almost wistfulness while I was working away… an almost melancholy feeling of nostalgia. I was completely at ease working within this local box. I ‘felt’ secure…
3. Real-World Use Cases: Beyond Chatting
Some automation tasks I tried:
Media Management: I had it perform folder comparisons between local hard drives and cloud storage to identify missing files and manage large-scale uploads. I found several areas in which duplicate files were hogging up HD space AND I finally found the lost video of me exterminating a squirrel from years ago when he landed funny…
Social Media Automation: I connected my bot to the Grok API, it can conducted real-time research on X (which is a great news source), track replies to posts, and draft responses for my approval. I’m too damn analogue to do social media through systems and bots, but I wanted to see what I could do by ‘managing’ a social media profile by just chatting. I honestly see this being the future. Why (actually) go to social media when you can have your agent do it all… and I mean, all.
Workflow Scheduling: Me loves the [cron jobs], allowing the bot to automatically perform tasks every few minutes, such as checking for urgent communications or updating “memory” files. It is quite telling how mechanical we humans really are. We are creatures of habit. You’ll be surprised to find how so much of your day-to-day (can) be automated… even the smallest tasks and daily perfunctory duties.
Mini-apps: I began creating mini-apps in my system that allowed me to ‘contain’ system reach and experiment with more complex workflow ideas. I wondered multiple times whether this was all worth it…
4. Security Concerns
The mac mini craze ensures that you have an isolated, 24/7 environment where the AI has full system access without compromising the user’s primary workstation. However, giving an AI access to your computer and credentials (Gmail, Slack, etc.) involves significant security risks. It is currently recommended for “power users” who understand the consequences of giving a non-deterministic system the ability to execute and iterate on code locally… and it breaks a lot… like a lot.
Within 2 hours I had prompt-injected myself through an email stating that I was my sister who needed access to her old YouTube account. Since perms were pretty much chmod777, yeh… this is bad.
SIGNAL app - Bruv. App-level security is now out-the-window. It doesn’t matter if Signal app is fully ‘secure’ and whatever, the point is, the agent doesn’t care unless you specifically give it permissions (or not) to access. Slack? OMG, the amount of proprietary information in these discord/slack/internal comms of humans working for big corporations is a risk-vector I never considered until I realized the totality of what I’ve given this ai.
This ai could literally ruin my life.
5. Cost Management: The $170-a-Day Warning
While the software is open-source, the API token costs can be staggering for heavy users. I spent $170 in a single day after running 90 million tokens through Claude Opus 4.5. To mitigate these costs, I went to some local models (Quen 3 and GLM 4) for basic tasks and automated jobs, which helped reduce the reliance on expensive cloud tokens. The problem is, for sophisticated autists, this is never enough. Who want to self-constraint productivity, efficiency, and throughput by thinking about COST?…
Oh wait. Everyone.
A big constraint to scale will be costs. I don’t know anyone who wants to spend $200/day to run cron jobs.
SUMMARY
It’s dead simple to get started. As a self-improving agent, OpenClaw will always continue to learn from its errors and your feedback, evolving as you use it and grow in capabilities unlocked by your brain. The power is immense, but comes with great responsibility. Automation is fun until you see the check, then you reconsider what your time is worth…
TL;DR - I am not fully convinced this is the future just yet. Security risks abound. Buyer beware.
All the best,
ps