github.devotd.
NewModel Context Protocol

goodnight000/api-anything — Turning websites into agent-ready APIs

goodnight000/api-anything is a local tool that learns inputs from the requests a website frontend sends, builds APIs that agents can call, and relearns to recover when the site changes.

Repositorygoodnight000/api-anything ↗

goodnight000/api-anything is a local tool that learns inputs from the requests a website frontend sends, builds APIs that agents can call, and relearns to recover when the site changes.

GitHub description: Turn any website into an API your agent can call. Learns from the site's own frontend, heals itself when the site changes.

stars
★ 1.3K
Forks
131
Language
TypeScript
License
MIT
Read original source ↗View on GitHub ↗

goodnight000/api-anything is a local tool that turns websites into APIs agents can call. It compares the requests a site builds and sends itself against sample inputs, stores only the input part as a template, and lets the generated API be called over MCP, CLI, and TypeScript to receive JSON. When a site changes and the requests stop working, it relearns and recovers.

The core idea is to observe the requests a frontend sends, generate the scraper automatically, and keep it as operations (the unit of work of an API) so agents get JSON without opening a browser. Calls are handled mainly over plain HTTP, and when a site blocks that, an alternative path through the browser is used. The repository ships reproducible scripts under bench/ (engine.mjs, import.mjs, transport.mjs, ground-truth.mjs), the raw material in results/, and a snapshot-based ground-truth check, so it forms a self-verifying structure. It is written in TypeScript, follows the MIT license, and the repository created on September 27, 2026 has now collected 1283 stars.

That said, the features included up front are limited to learned operations, write operations are not provided by default, and login sessions are managed by the tool itself. That it avoids browser memory cost is stated explicitly. The 1.5~2x speed advantage figure is a self-measurement on a single Mac environment, so no independent reproduction exists yet. Learning from the frontend to build agent-ready APIs is still at an early stage even in the MCP ecosystem. The self-verifying structure with reproducible scripts is in place, but independent validation backing the performance advantage is not.