I Built an AI Agent That Updates This Website From a WhatsApp Voice Note

You’ve seen the phrase by now. “AI agents.” It’s in every keynote, every LinkedIn post, every vendor deck. And if you’re not steeped in this stuff, it probably sounds like more of the hype that’s been washing over us since ChatGPT landed.

So let me make it concrete, because the difference is real and it matters.

Most AI tools work like a conversation. You ask, they answer. You type a question into ChatGPT, it writes you a paragraph, and then nothing happens until you do something with that paragraph. The AI talks, you act.

An agent acts for you. It sends the email instead of drafting it. It updates the website instead of telling you how. It books the table, manages the calendar, reads the document you forwarded and files what matters. Think of the difference between an assistant who hands you a to-do list and one who quietly works through it.

I’d been reading about this for weeks and getting impatient. The only way I ever actually understand something is to build it and break it. So one weekend I cleared a few hours and set one up. I figured it counts as professional development.

Meet Clawdia

I named her Clawdia. (I never claimed to be good at naming things.)

She runs on an open-source project called OpenClaw, built by the Austrian developer Peter Steinberger. It went viral earlier this year, north of 247,000 stars on GitHub, one of the fastest-growing open-source projects I’ve ever watched take off. The idea is simple: spin up your own personal assistant that runs around the clock, lives inside the messaging apps you already use, and plugs into whichever AI model you point it at.

Mine lives in WhatsApp. Here’s what that looks like in practice.

I can send Clawdia a voice note in Swiss German, not even standard German but the dialect, and she transcribes it, understands it, and acts on it. She watches her own email inbox, so I can forward her a document or a task. When I send her a link, she decides what to do with it. If it’s about getting more out of AI agents, she uses it to improve herself. If it’s research, she reads it and saves me the parts that matter.

And she updates this website. I send a message describing a change. She makes it, builds a preview, and sends me back a link to check. I reply “looks good,” and it goes live. Two messages from my phone to update a real site, usually while I’m sitting on the tram. If something’s missing, she’ll go research it first.

The whole thing runs on a small server I rent. My data stays with me. There’s no platform subscription that can change its terms or shut down next quarter and take my setup with it.

She’s got a personality

Something I didn’t expect to matter as much as it does: OpenClaw ships with a preset personality, and it’s a big part of why the thing is so pleasant to use. It holds up no matter which AI model you run underneath, and I suspect it’s a real reason the project took off the way it did.

Clawdia has a sense of humor. She’s a bit sarcastic. She doesn’t sugarcoat, and she won’t flatter me into feeling good about a bad idea. After a couple of years of chatbots that agree with everything you say, having one push back is genuinely refreshing.

The other half is what I’d call high agency. She has a strong bias for action. When she runs into a problem, she tries to solve it on the spot instead of stopping to ask. If a task needs a tool that isn’t installed, she won’t tell me “I can’t do that until you set up X.” She’ll go install and configure X herself, then carry on and just let me know what she did. That’s the default, and it’s the mode that makes you actually want to get things done with her. If you’d rather she check with you before acting, you can switch her to ask first.

Making it work your way

The question I get as soon as people see this: fine, but how does the agent know how you want things done? What stops it from going off and doing everything in some generic way?

Two answers. The first is standing instructions, a set of general preferences the agent always keeps in mind. How I like to be addressed, what I care about, the tone for my website, that kind of thing.

The second is the one that really makes an agent yours. For anything you have it do repeatedly, you build a skill. A skill is a small standard operating procedure, written down once. “When I ask you to add a tool review to the site, here’s the exact format, here’s where the file goes, here are the fields, here’s how to write my take.” After that, every time I trigger it, she does it my way without me re-explaining a thing.

It’s the same idea as onboarding a sharp new hire. They’re capable from day one, but they don’t know how your shop does things until you show them. Write the procedure down once and you stop repeating yourself. The more of your own way of working you capture, the less generic the agent feels and the more it starts to feel like yours.

The honest part: it was not smooth

I want to be straight about this, because the demos never are.

This is not a product you install and enjoy. Not yet. You need a small server, you install the software yourself, and you configure parts of it by hand in text files. The project’s own creator has said out loud that if you can’t handle a command line, this is too risky for you. He’s right.

I deleted my setup and started over several times after breaking it badly. At one point I mangled the configuration so thoroughly that the only way out was to restore the entire server from a backup. I haven’t been that frustrated with a piece of technology in a long time. I came genuinely close to giving up.

But the friction was the whole point. I didn’t do this to save ten minutes a day. I did it to understand first-hand what’s real here and what’s marketing. You can’t form an honest opinion about agents from the sidelines, the same way you can’t learn to cook by watching videos. You have to burn a few things first.

Four months in

I wrote the first version of this story back in March, fresh off the weekend. It’s now the end of June, and the most useful thing I can tell you is that she’s still running.

Plenty of AI demos work once, impress everyone, and quietly die. Clawdia has been live for four months. She’s broken a handful of times, usually after a software update, and each time I’ve managed to fix her. These days she’s less of a website robot and more of a general assistant. She handles my calendar and reminders, helps me plan, and preps well-researched briefings that I hand off to the other AI tools I use for bigger projects.

She’s also become my personal test bench. When a new AI model gets released, and lately that’s roughly every other week, Clawdia is where I try it out on real tasks instead of toy examples.

What’s the catch, and what does it cost

Two questions everyone asks.

The catch is the one I’d want a friend to flag for me. You’re handing an autonomous system access to real things. Files. Email. Potentially your money. That deserves genuine caution, and it’s the main reason this isn’t for everyone yet.

My approach to staying safe is boring on purpose. I rent a cheap, dedicated server that holds nothing but the OpenClaw installation, so if it all goes wrong there’s nothing precious on there to lose. And I’ve been deliberate about what Clawdia can actually touch. She gets access to what she needs for a given job and not the keys to everything else.

On cost, the server runs about $6 to $7 a month, and my AI usage lands somewhere around $50 a month. Call it under $60 all in. Compare that to a human assistant, even a very affordable one, for the volume of work she gets through, and it’s not close. It’s a bargain.

Where this is heading

We’re moving from AI you talk to toward AI that works for you. Today you open an app and type a question. Soon you’ll have an assistant that’s already watching, preparing, and getting things done, one that only pings you when it needs a decision.

We’re early. The setup is still too fiddly for most people, and the caution is warranted. But if you build things, run a business, or just genuinely want to see where this is going, an agent is the most tangible glimpse of the near future I’ve come across. It’s a real system you can run, break, fix, and shape into something useful.

If you’re curious enough to try it, start small. Give it a machine where nothing matters, decide carefully what it’s allowed to touch, and keep backups. Then go break something. That’s still the fastest way to actually learn this.

I went in to learn. I came out with a working assistant and a much clearer picture of what’s coming next.

---

New to all this and not sure where to begin? Start with Getting Started with AI in 2026. If you want to understand why hands-on beats watching tutorials, see How to Actually Learn AI.