July 9, 2026
Bots Behaving Browserly
Show HN: Reverse-engineering web apps into agent tools
AI sneaks into your apps, and the comments want it in Teslas and banks next
TLDR: A startup says its AI can learn a website’s hidden actions and turn them into tools, letting assistants do real tasks without needing special app support. Commenters were impressed but immediately pushed for riskier real-world demos like banking and Tesla, while one accidental-looking auto-post became the thread’s best joke.
A tiny Hacker News thread just served up a very 2026 kind of plot twist: a team says it built an AI helper that can sit inside a logged-in website, watch how that site talks to itself behind the scenes, and then turn those actions into usable tools for an assistant. In plain English, the pitch is: instead of clumsy screen-clicking bots or painfully manual setup, this thing learns the app’s hidden moves and starts doing useful jobs fast. The demos name-drop familiar favorites like Jira, Spotify, and even Hacker News itself, which gives the whole launch an extra wink.
And the community mood? Instant hype, with a side of “okay, but show me the scary version.” One commenter simply dropped a cheerful “so cool!”, which pretty much captures the first-wave reaction: impressed, curious, and ready to believe. But the spicier take came from the crowd member asking for demos on Tesla accounts, online banking, and movie ticketing—basically, “Cute demo, now throw it at the stuff people actually panic about.” That’s where the tension lives: everyone loves the magic trick, but they also want proof it works on messy, personal, high-stakes websites.
Then came the comic relief. A random “Testing automated HN posting from Hermes. Please ignore.” landed like a perfectly timed glitch in the matrix. In a thread about automation invading web apps, the funniest cameo was… an automated-looking post asking to be ignored. You couldn’t script a better punchline, except apparently now maybe you can.
Key Points
- •The product is a browser-based agent that runs inside authenticated web apps and observes internal API calls to generate reusable tools for AI agents.
- •The system is presented as an auto-generated MCP-like server that updates tool definitions as the target application changes.
- •The post includes demos using Jira, Spotify, Hacker News, and a full end-to-end demonstration.
- •Each generated "recipe" includes endpoint and method, authentication details, response schema, input schema, and a human-readable tool description.
- •The author says this approach avoids modifying source code, reduces dependence on brittle browser automation, and can execute actions directly within existing applications.