I've started playing with OpenClaw

I've spent a few weeks tinkering with OpenClaw, the agent half of YouTube is talking about. This isn't a definitive review, just the first impressions of someone who has installed it, given it jobs, watched it break other things and had to cut back its permissions. I'll keep updating it as I go.

If you've opened YouTube in recent months, you'll have come across some video about OpenClaw with an astonished-face thumbnail: "the assistant that runs your whole life", "fire your secretary", "I left it alone for a week and look what it did". With that soundtrack, you'd think all you have to do is install it and sit back and watch.

The real story is shorter and more curious. It was born in 2025 under the name Clawdbot, was renamed Moltbot and in January 2026 took the name OpenClaw. Within a few weeks it became one of the most popular projects on GitHub, the platform where a good chunk of open-source software is published. Three names in a few months already hints at the pace of all this.

I tried it out of pure professional curiosity: several clients asked me about it in the same week and I didn't want to answer with what I'd seen in videos. I work with agents every day and it was time to find out what this one brings that the others don't.

A warning before going on: these are my first weeks. What I describe here may change, and I'll update it, marking what has changed.

OpenClaw is a free, open-source agent you install on your own machine, connect to the AI model of your choice (paid or local) and control from messaging apps such as Telegram, WhatsApp, Slack or Discord. Instead of just answering, it does things: reads and writes files, runs commands, browses websites or sends messages.

If you're not sure what an agent is, the most useful difference is this: the chat tells you how to do it and the agent does it. You give it a goal, it decides the steps, uses tools and hands you back the result. That makes it useful and also delicate, because if it gets something wrong the mistake doesn't stay in an answer, it stays in your files.

OpenClaw doesn't come with its own model. What it provides is everything else: the tools, the memory and the connection to messaging. That "everything else" is called a harness. The model talks; the harness gives it hands.

I installed it on one of my six VPSs, a small one I was underusing and have left dedicated to it alone: if something goes wrong, I can wipe it and rebuild it in twenty minutes. The reason is simple and non-negotiable: never on the computer where I keep what matters. A program that can run commands on a machine can run the wrong command on that machine, and I'd rather that happened where there's nothing to lose.

As the model I connected Claude through the API, which is the door through which a program uses the model, paying per use. I also tried a local model on one of my internal servers, but for this kind of work, with many chained steps, it got lost far too easily, so I left that for later.

Installation took me about four hours, spread over two afternoons, and not because of the program itself. The stumbles were an old version of Node.js on the server that had to be updated, the Telegram bot, which didn't respond because I'd copied the token (the key that identifies the bot) wrong and a firewall I myself had left too tightly closed. Nothing serious, but nothing like "install it in five minutes" either.

I gave it three specific jobs, all from my phone via Telegram:

1. Check three supplier websites every morning and let me know if any price on a list of products had changed.
2. Tidy a folder of test documents, renaming them by date and type.
3. Prepare a daily summary of the server's own logs and warn me if it saw anything odd.

The one it did best was the first. For ten days in a row it sent me a message at eight with the changes, and the day a supplier raised the price of a product it caught it and sent me the link. Doing it by hand used to take me about fifteen minutes a day, which I'd stopped doing anyway because it bored me. That was the first time I felt what a real agent promises: something that works while you do something else.

The second went well on the third attempt: the first two times it invented a different date format for each file. When I gave it an exact example of how I wanted the names, it nailed it. The third I'll tell in the next section, because that's where the crash came.

The logs task looked like the most innocent. I asked it to check the server logs and tell me about anything odd. It found a service throwing errors and, instead of telling me, decided to fix it: it changed a configuration file and restarted the service. The service didn't come back up. I found out because the next day the eight o'clock message didn't arrive. Luckily it was on its own separate VPS, with nothing else on it; I fixed it by restoring the configuration from a backup in about half an hour.

The permissions scare was another thing. Going over what it had done, I saw that in one of its attempts to tidy the folder it had also listed the user's entire home directory, because it had permission to and thought it was a good idea. It didn't delete anything, but it could have. At that moment I understood the problem wasn't the agent but me, for having given it permission for everything.

What it cost me: about twelve hours between installation, tests and fixes and about €35 of model usage in three weeks, mostly because of the logs task, which read a huge amount of text every day.

Afterwards I cut back permissions: a dedicated system user for the agent, with access to only one working folder, no restarting services and confirmation by message before any command that changes something. Now it's slower and more tedious. I also sleep better.

In my day to day I work mainly with Claude Code and other AI agents spread across my servers and my VPS. They're agents I sit down to give a specific job, watch them work and close the session when they finish.

What OpenClaw brings is something else: it's always on, remembers what it has done from one day to the next and I can write to it from my phone as if it were a colleague. That combination, messaging plus memory plus always being there, I didn't have set up in such a comfortable way. What it doesn't bring is more intelligence: the brain is still whatever model you connect, and for delicate code work I still prefer Claude Code, where I see every step.

In my setup I'm keeping it for periodic monitoring and alerts: things that need looking at every day and aren't worth me opening the computer for.

With n8n, the block-based automation tool I used before, the usual thing is happening: I use it less and less because agents are replacing it. I had the price monitoring half built in n8n and abandoned it when I saw OpenClaw solved it with one sentence.

If you're up for it, these are the rules I'd set from day one, not from the first scare:

• Install it on a separate machine or a test VPS, never on your work computer.
• Give it minimal permissions: one folder, its own user and nothing more until you know how it behaves.
• No client data or important accounts while you're testing.
• Keep an eye on model spending; if you connect it to a paid one, set a limit on the account.

Handing a program your house keys calls for the same caution as handing them to a person. What has happened to me with agents in general, from permissions to costs, I tell in The problems of working with agents.

This is a provisional verdict, with the test started on 1 September 2026 and written a few weeks later.

What's true in the promise? There's something real: an always-on agent you talk to from your phone, doing repetitive tasks while you get on with life, works and makes a difference. What's noise? The "runs your whole life" bit. For me it has run the monitoring of a few prices and broken a service along the way.

Who is it for? Someone who wants to really understand what an agent is, with its hands and its dangers, and has a separate machine to play with. Who isn't it for? Someone taking their first steps with AI: it's not the most comfortable place to start, and the installation alone weeds out quite a few.

I commit to coming back to this article after a few months with it and telling you whether it's still in my setup or has ended up in the folder of forgotten experiments.