Teach an agent a website once

Teach an agent a website once, then let it call the site like an API. That's API Anything's pitch, and the README backs it with an honest benchmark.

goodnight000/api-anything is MIT licensed, created in late September, with about 1.2k stars. It watches a site's own page make a request, works out which parts are your inputs, and saves that as an operation. Your agent then calls the operation over MCP, a CLI or TypeScript and gets JSON back over plain HTTP, with no browser.

The README includes a head-to-head: the same agent (Claude Opus 5.5) on Google Flights tasks, once driving headless Chrome through Playwright and once calling the saved operation. Over 5 runs each, the API route answered 1.5 to 2x sooner and read about half as many tokens, at the same cost or less. On the one-date task the medians were 12.9 s and $0.166 for the browser against 8.4 s and $0.074. All 20 answers were correct.

Read the limits too: one site, two tasks, five runs each, run by the author. And the package isn't on npm yet, so you install it from the repo.

What I like most is the failure behavior. A failed call returns a reason and a next step, not a blank page.

Which website would you teach an agent first?