I turned Markdown into a protocol for generative UI (fabian-kuebler.com)

130 points by FabianCarbonara ↗ HN
There's a lot of work happening around both generative UI and code execution for AI agents. I kept wondering: how do you bring them together into a fully featured architecture? I built a prototype:

- Markdown as protocol — one stream carrying text, executable code, and data

- Streaming execution — code fences execute statement by statement as they stream in

- A mount() primitive — the agent creates React UIs with full data flow between client, server, and LLM

Let me know what you think!

35 comments

[ 2.8 ms ] story [ 63.1 ms ] thread
(comment deleted)
If you're still looking for a name let me suggest "hyper text".

It embodies the whole idea of having data, code and presentation at the same place.

If you're open for contributions I already have an idea for cascading styles system in mind.

> If you're still looking for a name let me suggest "hyper text".

Perhaps "WWW SPA document"? Using markdown with highly-progressive fenced blocks?

Hypertext (one word, coined 1960s) is quite a broad category. Subcategory "WWW" could fit, as TFA seems WWW-ish. A markdown document format, and progressive rendering of tags and code, seems HTML-like. Though with greater progressiveness - code blocks with streamed execution rather than merely compilation. The progressive JSON callbacks, React, integrated client and server code execution, and server-side rendering, seem closer to WWW SPA than to HTML. Though SPA files often seem more "source" than "document". And the multiple-page "App"-ness of SPA doesn't fit well. SPA seems a better fit than "full-stack". Perhaps some name analogous to "isomorphic javascript"...?

Or more precisely, isn't this reinventing notebooks (not the first JS-centric notebook either)?
There’s definitely a lot of merit to this idea, and the gifs in the article look impressive. My strong opinion is that there’s a lot more to (good) UIs than what an LLM will ever be able to bring (happy to be proven wrong in a few years…), but for utilitarian and on-the-fly UIs there’s definitely a lot of promise
(comment deleted)
OpenUI and JSON-render are some other players in this space.

I’m building an agentic commerce chat that uses MCP-UI and want to start using these new implementations instead of MCP-UI but can’t wrap my head around how button on click and actions work? MCP-UI allows onClick events to work since you’re “hard coding” the UI from the get-go vs relying on AI generating undertemistic JSON and turning that into UI that might be different on every use.

The nice thing about standards is that you have so many to choose from
I quite like this! I've been incrementally building similar tooling for a project I've been working on, and I really appreciate the ideas here.

I think the key decision for someone implementing a flexible UI system like this is the required level of expressiveness. To me, the chief problem with having agents build custom html pages (as another comment suggested) is far too unconstrained. I've been working with a system of pre-registered blocks and callbacks that are very constrained. I quite like this as a middleground, though it may still be too dynamic for my use case. Will explore a bit more!

In an agentic loop, the model can keep calling multiple tools for each specialized artifact (like how claude webapp renders HTML/SVG artifacts within a single turn). Models are already trained for this (tested this approach with qwen 3.5 27B and it was able to follow claude's lead from the previous turns).
The bots that read the instruction and yet add the emoji to the _beginning_ of the PR title though. Even bigger red flag I guess?
Very cool. I'm imagining using this with Claude Code, allowing it to wire this up to MCP or to CLI commands somehow and using that whole system as an interactive dashboard for administering a kubernetes cluster or something like that - and the hypothetical first feature request is to be able to "freeze" one of these UI snippets and save it as some sort of a "view" that I can access later. Use case: it happens to build a particularly convenient way to do a bunch of calls to kubectl, parse results and present them in some interactive way - and I'd like to reuse that same widget later without explaining/iterating on it again.
would be nice if it wasnt just ui but other form like voice narration, sounds ect
I see potential to take over Notion's / Obsidian's business here. Imagine highly customizable notebooks people can generate on the fly with the right kind of UI they need. Compared to fixed blocks in Notion
I will say I came upon this same design pattern to make all my chats into semantic Markdown that is backward compatible with markdown. I did:

````assistant

<Short Summary title>

gemini/3.1-pro - 20260319T050611Z

Response from the assistant

````

with a similar block for tool calling This can be parsed semantically as part of the conversation but also is rendered as regular Markdown code block when needed

Helps me keep AI chats on the filesystem, as a valid document, but also add some more semantic meaning atop of Markdown

> AI chats as a valid document

So many formats, with different tradeoffs around readable/parsable/comments/etc. I wish there was a "universal" converter. With LLM's sometimes used to edit chat traces, I'd like ingestion from md/yaml, not merely a "render from message json".

So .json `[{"role": "user", "content": "Hi"}` <-> .md ` ```json\n[{"role": "user", "content": "Hi"}` <-> above ` ```user\nHi` <-> `# User\nHi` <-> ` ```chatML\n<|user|>\nHi` <-> .html rendered .md, but with elements like <think> and <file> escaped... etc.

A2UI is Google's take — declarative JSON, tool-calling based, predefined component catalog. Clean and safe but constrained.

My approach is the opposite bet: full code execution instead of tool calls. The agent can build any React UI from scratch with the full power of code — including client-server data flow, callbacks, streaming data.

Brainstorming, perhaps `<<named-block-code-transclusion>>`? It goes against the grain of "eval() line-by-line", even if it's handled ASAP. But it might relax the order constraint on codegen. Especially if the UI gets complex, or rendered on a "pane off to the side".
Interesting idea! The slots mechanism already handles some of this — you can mount a skeleton first and fill in named sections later as the LLM generates them. But true out-of-order transclusion could be useful for more complex layouts. Worth exploring!
Oops - that's not transclusion. Merely literate programming tangles.