Frigade

I made Jev beat all of Pokemon Red for less than 2 dollars with a median decision time of 400ms. While completing the game took a bit longer than I expected (much of it spent goofing around in a cave), I have no doubt these low-latency decision models will change everything.

We gave Jira, Spotify, and Airbnb an AI agent with 100+ built-in Skills, and we never touched a line of their code. Every "AI copilot inside your app" the past two years has the same ceiling: it can answer questions about the product, but it doesn't use or even understand the product. The fact is most apps still don't have an MCP server or agent-ready API to plug into, and driving the UI with computer-use is too slow and brittle to ship. Skills by Frigade solves this. Our agent uses an app like a real user, watches how the app calls its own authenticated APIs, and then turns those calls into tools it can reliably run. Ask it for something and it (securely) calls the real endpoint on the user's behalf instead of clicking through screens or paraphrasing a help article. It works the same way on any web app, including yours. Frigade Skills are now live. Full demo in the comments 👇

We just moved the entire Frigade marketing site off our no-code CMS back to @nextjs in a week. We're not using any external services other than @vercel. All content is stored i markdown files so not even any need for a headless CMS. Using markdown also allows to easily leverage Claude Code for authoring and cleaning up content and keeping a consistent voice and tone. The whole pitch of no-code was solving one problem: engineering attention is too expensive to spend on a marketing site, so route around it with a tool a non-engineer can drive. That problem is mostly gone now. Tbh it turned out to be surprisingly easy. And for something as straightforward as a marketing site, the stakes of vibe-coding are way lower than on a production app. We have no issue having non-technical teammates push to main on that repo. My bet is we're not the only team that's going to make this trip in the coming months.

Hardest part of building AI products is that nobody believes you until they see it. You can join whatever hype circle is happening on X and talk about how productive AI has made you. But there are so few real raw demos out there. So here's one. I recorded Frigade working inside @vercel. Not scripted or staged. The AI learns the product, highlights UI elements, and walks users through actual workflows. When the product changes, it adapts. I've been building this for over a year and it still blows my mind. Easier to just watch it.

The market has been flooded with AI chatbots, and most of them have the same problem: They are isolated from the product and the user experience. So when a user gets stuck, they have to explain everything to the bot from scratch. It's an incredibly clunky experience trying to describe software problems in plain words. And then because of that, the output is just not that helpful. Usually a wall of bullet points copied directly from the help center. We ran a simple test: same scenario, same user problem. One assistant had workflow context. The other didn't. The usefulness wasn't even close.