Magnitude — браузерный агент на визуальном ИИ

★ 4.1K

Magnitude is an open-source TypeScript library that lets you control a browser in natural language: the agent sees the page as an image, uses a visually grounded model to pick pixel coordinates and performs precise mouse and keyboard actions. Reach for it when you need to automate tasks on sites without an API, when you want to test a web app in plain language with visual assertions, when you need structured data from a dynamic interface, or when browser checks must run in CI/CD. Inside: act (high-level tasks and granular actions, with optional data passed to the agent), extract (data matched to a Zod schema) and a built-in test runner, magnitude-test. Start with npx create-magnitude-app, or npm i --save-dev magnitude-test && npx magnitude init. It needs a large visually grounded model: Claude Sonnet 4 is recommended and Qwen-2.5VL 72B is compatible. Licensed under Apache 2.0. The vision-first approach relies on coordinates rather than numbered boxes drawn around DOM elements, so it does not depend on page markup. The direction is browser control and web automation. Suits TypeScript projects with access to a compatible model; the README marks deterministic caching as still in progress.