Currently LLMs can only stream raw text and users can only respond in raw text.
This open, MIT license standard allows LLMs to send simple, secure, JSON code blocks which get rendered into widgets.
The widgets can emit events, which your application can capture and send back as text to the LLM.
Currently I've coded a proof of concept using an extended marked.js parser and Svelte 5 renderer, but any parser or renderer (Remark, React etc.) can support this standard.
Please let me know what you think!
