What I built, and learned at my first OpenClaw hackathon

Going in solo

When the hackathon started I thought it'd be a great opportunity to test myself as a vibe coder, to see what I could bring to the table in 24 hours. I wanted to go solo on this one. I didn't know enough about OpenClaw to confidently contribute to a team, and the last thing I wanted was to be dead weight.

The original idea

I work in the financial wellness space and know how much people want to save money on small things that add up every month. On average, people in the US spend around $200 a year on subscriptions they don't use. Subscriptions felt like the right problem. Every other app just helps you track them. Except that one app (🚀💰) that makes you fill a form while humans do the cancellation in the background.

I wanted to build something smarter. An AI agent that actually understands your usage and tells you whether a subscription is worth renewing. And if you wanted to cancel, it does it for you.

The agent was going to live inside Neo, a browser built by Gen, powered by OpenClaw.

The reality check

I had never worked with OpenClaw before. I thought I could figure it out as I went, the way I usually do with React or SwiftUI. That assumption didn't last long.

The first few hours were just trying to understand what I was looking at. Then trying to add skills through the web UI. I tried installing a browser through the terminal. Gateways failing, commands not going through, no real sense of what I was even doing. This wasn't like vibe coding a component.

I gave up around the 7-hour mark. The last thing I needed in my life was more stress. I took a 4-hour power nap to compensate for the 11-hour sleep debt I had according to my Oura ring.

The pivot

But for some reason it kept bothering me. I work on a lot of side projects but I never finish. This felt like another one slipping away.

So I sat back down with one goal. Just get OpenClaw to open a website on its own browser.

About 15 hours into the hackathon I realized I might need SSH access, something they'd only give you if you specifically asked for it. Once I had it, I came across something called Playwright, installed it through the terminal, pointed it at amazon.com, and got a screenshot back.

I shelved Plaid. It was the piece that would've made subscription tagging accurate and usage reports timely. Instead I started building piece by piece: an extension that maps your browsing against a hardcoded list of subscription services.

There were still port issues, tunnel stability problems, constant switching between the remote instance and my local machine.

I named it ClawCancel

The browser extension is the channel. You get a notification and a report on which subscriptions you're actually using, what you're wasting, and whether they're worth keeping. Also sent to you via a Telegram bot. You can also chat with the AI in the extension to dig into your usage. And if you want to cancel something, one command triggers it. OpenClaw's browser relay extension runs the action on your local browser, since headless browsers kept getting flagged as bots by websites.

Submitting it

With 2 hours left for the deadline, I submitted my code and demo video. I felt the satisfaction of completing something. I stopped chasing perfect and just built something that worked. For a 24-hour hackathon with a steep learning curve, that turned out to be the right call. Not everything made it in. The Plaid integration, the part that would've made it genuinely smart, never got built.

Takeaways

I walked into unknown territory and decided to figure it out on the way. That works sometimes. It does not work at a hackathon.

For someone like me going into unknown territory, I'd recommend going in with at least a rough plan so you can scope out things that aren't worth chasing if you're short on time. I dropped Plaid not because it was too hard but because there wasn't enough time to do it right, and knowing that earlier would've saved hours. Plan mode exists in Claude and other AI tools for a reason.

In the end, no matter what, finish and submit it. Even if it just works on your local machine. Btw, mine actually worked on different machines.