Turning web apps into agent tools via API reverse-engineering

Instead of the brittle "computer-use" approach where an LLM just clicks buttons like a human, this method intercepts the actual network requests. It constructs what they call a "recipe," which is a structured package containing:

The technical advantage here is the self-healing aspect. Since the agent is essentially "watching" the app, if the underlying API schema changes, the tool updates itself. This solves the maintenance nightmare of traditional integration work. I noticed they specifically mentioned that GraphQL was a massive headache for standardizing these recipes, which makes sense given how dynamic those queries can be.

It's a much more efficient way to give an LLM agency without burning massive amounts of tokens on visual reasoning. You can see how it handles different environments in these specific demos:

https://demo.frigade.com/hn?skill=jira
https://demo.frigade.com/hn?skill=spotify
https://demo.frigade.com/hn?skill=hackernews
https://demo.frigade.com/hn?skill=full-demo

From a security standpoint, it's a cleaner architecture because the agent acts directly within the existing authenticated session rather than needing a complex third-party proxy. It's a solid step toward making AI actually useful in enterprise workflows where APIs are often too complex for a standard RAG setup to navigate.