My Weekend with OpenClaw

TLDR: I installed Linux Mint on an old notebook, set up OpenClaw, and lived with an AI agent for a weekend. It searched through my notes, reminded me of my own principles, and pointed out that my son had a fever for two days. Here is what I learned, and why I still turned it off for now.

FOMO Got Me

I admit it: FOMO got me. Everywhere you read about AI agents that don’t just answer, but act. That create files, install software, check appointments. No longer just a chat window, but a digital assistant that runs on your own computer and does things.

The topic wasn’t entirely new to me. I had already set up my blog with Google Antigravity and gained my first experiences with AI in the terminal. That was impressive, but still conformed to expectations.

OpenClaw is a different beast. An open-source framework for personal AI agents. The agent runs locally on your machine, has access to your file system, can research the internet, and you can reach it via Telegram, a web interface, or directly in the terminal. When I read about it, I wanted to try it out. So I cleared my schedule for a weekend.

How I Went About It

An Old Notebook and a USB Stick

Step one: Hardware. I still had an old notebook lying around that had been gathering dust for years. Perfect for an experiment where I couldn’t break anything.

I needed Linux on it. Which distribution? For once, I asked Grok on X. Normally I rarely use Grok, but for a project like OpenClaw, which is only a few weeks old, real-time knowledge from the X community seemed more relevant to me than the training data of a normal LLM. The recommendation: Linux Mint. Windows-like, beginner-friendly, stable. Grok also explained right away how to make the USB stick bootable.

The installation itself was unspectacular. Insert USB stick, F12 on boot, start from the stick, click “Install”. Erased old hard drive, put new system on it. Then the usual Linux onboarding: customize desktop, security settings, configure firewall. Nothing you can’t manage with a little patience and a search engine.

Installing OpenClaw

The installation of OpenClaw itself followed the official guidelines. It didn’t go completely smoothly; a few missing libraries and dependencies had to be added first. For example, I didn’t know which package manager to use to load certain dependencies. But this is exactly where the advantage of already working with AI in the terminal shows: I simply gave the error messages to Claude and solved them step by step. Not glamorous, but it works.

Then the onboarding wizard. Here you choose your model. I opted for OpenRouter: one API, one billing, access to all relevant models. As the concrete model, I chose Claude Opus 4.6, the best (and most expensive) available. My logic: The first time around, I want to see the best quality. I can optimize and save money later when I know what’s possible.

There were also a few errors during the onboarding itself. Nothing dramatic, but you should be prepared that you don’t just click “Next, Next, Finish”. More like: read error message, google or ask the AI, fix, continue.

Personalizing the Bot

The last step before starting: setting up the bot. I gave him the name “Karl”. The configuration parameters are set in natural language; no YAML tinkering, you simply describe how the agent should behave.

I chose Telegram as the gateway. Easy to set up, Telegram has good bot support, and I can reach the agent from my phone. WhatsApp would have been the alternative, but then I would have needed either a separate number or linked my private number, which would have given the bot access to my contacts. Not an option.

Three Ways to Talk to Karl

After setup, I had three ways to interact with my agent:

Terminal UI (TUI): Right in the command line. Text in, text out. Has a hacker vibe that I honestly liked the best. You see exactly what the agent is doing, what commands it’s executing, which files it’s touching. I also downloaded Ghostyy as a modern terminal solution for this.

Web Interface: Runs on localhost in the browser. More configuration options, clearer for longer conversations. Handy if you want to tweak settings.

Telegram: The game-changer for everyday life. I can send Karl a message from the couch. Or record a voice message. I no longer have to sit at the computer to use my agent. Everything that came after the initial setup, I did mostly via Telegram.

The Aha Moments

Chatting with My Notes

My first use case: I wanted Karl to be able to read, expand, and create my Obsidian notes. My notes are normally kept in Obsidian and synced between my devices via Syncthing. The notebook was freshly set up, so none of that was installed.

So I told Karl what I needed. And then something happened that actually felt like magic: Karl independently installed Obsidian and Syncthing, did the configuration, configured the keys between the devices, and set everything up. I just watched, provided the keys, and occasionally gave confirmation.

That was the moment I understood what distinguishes an agent from a chatbot. A chatbot would have explained to me how to set up Syncthing. Karl simply did it.

Intelligent Finding

As soon as Karl had access to my notes, I could ask him questions like “Where did I write down my thoughts on Deep Work again?” without knowing the exact file name. Once I explained the folder structure to him, he found the right files reliably. No grep, no manual searching. Just ask.

Goodbye Inbox

My previous note-taking system looked like this: during the day, I took notes via the iPhone Action Button, which were then appended to an Apple Notes Inbox note. In the evening, or at the latest a day later, I processed the Inbox notes and moved them to their proper place in my Obsidian Vault. Since these aren’t just thoughts but also small tasks to execute, this easily took 30–75 minutes every evening. An activity necessary for my personal sense of order, but it costs time.

Now I could work differently: I could tell my bot in Telegram my thoughts and notes, and he writes them in the right place with the correct tags and backlinks for me. Additionally, he can execute smaller tasks. Cancel a subscription here, look something up there, etc.

Accelerated Workflows

I can automate smaller workflows with OpenClaw. For example, my blogging workflow used to look like this: dictate my blog article with Apple Voice Memos while on a walk. Then transcribe it at home with Aiko (OpenAI Whisper) and have Claude with Opus 4.6 formulate it into a blog article. Then copy-and-paste the text into my Obsidian as a draft, and manually edit and finalize it there. Then duplicate this final draft for my blog in Antigravity, where I host my blog. There, Gemini adds further metadata and generates a header teaser image.

Doesn’t take forever, but it’s a lot of media breaks. My new workflow can now look like this: I dictate my blog article in Telegram for Karl. He automatically files it in the right place in my Obsidian Vault as a Markdown file. At home, I can do my manual final editing work. Afterwards, I tell Karl to upload the post to my blog GitHub repo, and I just let him handle the smaller metadata and image generation tasks that I would otherwise have done in a separate tool (Antigravity).

So this is already largely automated, starting from my voice dictation on the walk. Furthermore, over time Karl learns my preferences, which he saves in his MEMORY.md. In the manual process, these would otherwise always have to be implemented manually over and over again.

Proactive Reminders

This use case surprised me the most. I have a collection of core principles and self-instructions in my notes. About 150 of them. Things I wrote down at some point because they were important to me, and which are supposed to influence my actions from then on. The problem: I never look at them. The notes exist, but they don’t take effect.

So I asked Karl if he could randomly pick 5 of my core principles daily via a heartbeat—a cronjob that runs every 30 to 60 minutes—and discuss them with me. Not preachy, but as a conversation starter. “Hey, you once wrote down that you always want to check the ABCs when making decisions. Did you do that last week with the X thing?”

It worked. Suddenly, notes that lay unread for years become alive again. They become actionable. That was a real aha moment.

Contextual Knowledge

The moment that impressed me the most came unexpectedly. My son had a fever, and I wanted to document that in the diary in the evening. I told Karl: “Please write in the diary that the little one has a fever of 38.7° today.”

Karl’s answer: “I see that you already noted a fever for him two days ago. That’s three days now. Maybe a doctor’s visit would be advisable.”

I had completely forgotten that. Karl hadn’t, because he has access to my diary and can make connections that slip my mind in the stress of everyday life. This is an agent that reads text files and puts two and two together. But the impact at the right moment is enormous.

The Thing About Security

I want to be honest here, because this is exactly the point where the enthusiasm takes a hit.

An AI agent that can read my personal notes like a diary, health notes, or professional thoughts is a security risk. Period. The attack vectors are real: third-party skills can contain Markdown files with hidden instructions. Cheaper models are more susceptible to prompt injection. Websites that the agent researches can contain manipulated content that influences its behavior.

Opus 4.6 is comparatively robust against prompt injections; that’s one of the reasons why I started with the most expensive model. But “comparatively robust” is not “secure”. And as an end user, I can hardly fully oversee the attack vectors.

My decision after the weekend: I won’t let Karl run permanently. I use him consciously, now and then, to get a feel for his capabilities. And I’m waiting for further updates and hardening before I give him permanent access to my most personal data. The potential is huge, but the responsibility lies with me.

What It Costs

The core setup, i.e., installing Linux, setting up and configuring OpenClaw, took about two hours. After that, I spent a few more hours playing around via Telegram, trying out use cases, getting to know Karl.

Functionally, everything ran flawlessly with Opus 4.6. No errors, no crashes, consistently good results.

But: My 20-dollar limit at OpenRouter was already used up during the core installation. I had to increase it to 50 dollars, and that just barely sufficed. Opus 4.6 is expensive, and an agent that works independently consumes significantly more tokens than a simple chat.

The next step is clear: Set up a routing system. And that’s exactly what I did on Monday morning.

The Routing System

The idea is simple: Not every task needs the most expensive model. Karl remains the “mastermind” on Opus 4.6; he decides, plans, and coordinates. But for the actual work, he delegates to cheaper sub-agents. Setting it up was straightforward: I simply told Karl which models I want to use and when. He recorded this in his notes.

The result is a three-tier system:

• Opus 4.6 (€5/€25 per million tokens): Remains the main session. Complex decisions, planning, security questions—everything where judgment counts.
• Sonnet 4.6 (€3/€15): Writing tasks, blog articles, diary entries, summaries. Everything that needs quality but not full brainpower.
• Kimi K2.5 (€0.45/€2.20): Simple file operations, formatting, creating lists. Ten times cheaper than Opus with solid results.

In addition, I deposited Codex 5.2 for hard technical tasks and coding, as well as Gemini and MiniMax as on-demand options for testing.

Specifically: The diary entry from Sunday, which I dictated to Karl via voice message, ran on Sonnet instead of Opus. Had I run it on Kimi, it would have been ten times cheaper again. Perfectly sufficient for pure formatting and typing. This way, the daily consumption can realistically be pushed down to a fraction of the first experimental days.

Tip: What If You Don’t Have an Old Notebook?

Not everyone has an old ThinkPad in their desk drawer. Three alternatives:

• Mini PC: Amazon has usable devices starting at around 200 euros. Small, quiet, easily sufficient.
• Used Notebook: Starting at around 300 euros you get something solid. Doesn’t have to be a powerhouse.
• Virtual Machine: For 10 to 20 euros a month at a cloud provider. Has the advantage that everything is isolated, always online, and perfect for experimenting.

Conclusion

A weekend with an AI agent showed me where the journey is heading. The setup is not yet for people who flinch at the word “Terminal”. But anyone willing to invest a few hours and not be deterred by error messages will get something that feels fundamentally different from ChatGPT.

The agent doesn’t just think along, it acts. It installs software, finds connections in my notes, and reminds me of things I have forgotten myself. It’s not a finished product yet, and then there are the security issues. But the potential that I experienced (instead of read about) this weekend won’t let me go.

Karl mostly sleeps now. But every now and then I wake him up.