Ask an AI to help you buy something and you usually get an essay. Then you open ten tabs to check whether any of it is right.
My Hermes bot can now answer with something you use instead: cards you can compare at a glance, numbered sources you can click, and a slider that narrows the list without asking it again.
ChatGPT does something similar with GPT-6 (OpenAI calls it Intelligent UI). I wanted it in the agent I actually work with, the one with its own tools, memory and files, so I built it into Hermes on OpenClaw Launch.
It's live for every Hermes bot there, and it stays off until you switch it on.
What it looks like
In Quick Chat on the dashboard there's a chip next to "/ commands" called Interactive. I turned it on and asked the kind of question I normally lose an evening of tabs to:
"I want a robot vacuum that handles pet hair, budget up to $500. Search the web for 4 strong options you can buy right now, with their current price and battery life from what you find. Show them as cards with a photo of each (use real image URLs from your search), and let me filter by max price with a slider."
A few minutes later the answer was four cards, not an essay. A Roborock at $249.99, a Dreame at $269.99, and two eufy models at $339.99 and $499.99 with coupons. Each card had a product photo, the battery life, what it's good and bad at, and numbered sources I could click. A slider at the top cut the list down as I dragged it.
That's the part I care about. Normally I'd get a page of text and then open ten tabs to check it. Here the checking came with the answer. It even warned me that the battery numbers come from different cleaning modes, and that it hadn't tested the coupon prices at checkout. Dragging the slider costs nothing, because nothing goes back to the model.
It's one run on my own bot, checked October 9, so treat the prices as that day's prices.
How it works
The obvious way to do this is to let the model write a whole web page and load it in an iframe. That works, but it's a lot of code to trust. OpenAI describes their version as a library of native components plus a compiler, and I went the same direction.
The bot writes a small markup block in its reply. Tags that look like JSX, a little JavaScript for state, if and each blocks. The browser compiles it, runs the code inside a sandboxed frame and two isolated workers, and draws the result with a fixed set of parts:
- rows, columns, cards and tiles
- text, buttons, inputs, sliders, toggles and selects
- tables, plus line, area, bar and pie charts
- drawings, with simple animation for games and diagrams
- pictures, from the bot's own files or the web
- a look of its own: dark or light, an accent colour, gradients
- 246 icons and a copy button
The view's code can't reach the page, your cookies, storage or the network. Pictures are the one thing it can show from outside, and it doesn't fetch them itself. Our server fetches every picture written into the view as soon as it appears, so nothing you type or click inside it makes a request. If a view is broken, too big or stuck in a loop, you get the code instead of the view and the chat carries on.
Every view has a "View code" link. I wanted to see exactly what the bot wrote, and I figured other people would too.
Views can also draw and animate, show pictures, and pick their own look. The vacuum cards above use the pictures.
Ask it how a ballpoint pen works and it draws an exploded view on a dark card, with the cap, barrel, ink tube, tip and ball each labelled. I never asked for dark.
What it costs
My bots run on tokens I pay for, so this was the first thing I measured, on my own bot:
- a normal call there already sends about 22,000 input tokens (system prompt and tools)
- with Interactive on, the format note adds about 1,900 tokens to the first message, roughly 9% more input. It grew when drawings, pictures and styling went in, and I'm counting from its length
- after that most messages carry a 60 token reminder, and the full note goes again every 10th message
- over a long chat I'd expect 10 to 15% more input, and most of it is sent as cheap cache reads
- plain text replies cost no extra output
- the tip calculator I tested first was about 600 to 700 output tokens, nearly all of it the view's code. A text answer would have been 100 to 200
Turn the chip off and nothing gets added to new messages.
Where ChatGPT is ahead
Honestly, in a lot of places. It's on by default and the model decides by itself when an answer should be a UI. OpenAI says it trained GPT-6 to make that call. And it's rolling out to Free users too.
Mine is a switch you turn on, and it's only as good as the model writing the markup. So far I've tested it with GPT-6.1 Sol. I haven't tried weaker models yet.
Where a Hermes bot is different
The view comes from an agent that already has my context. It can run a web search first and put what it found into the view, sources and all, like the vacuum cards above. It has its own memory, skills and files on its own server. And the view is plain text in the reply, so it doesn't care which vendor's model writes it.
The same bot keeps answering on Telegram and Discord in plain text. Those apps can't draw a live view, so Interactive only lives in the web chat.
One caveat
Turning the switch off stops new notes. In a chat where it was already on, the earlier notes are still in the history, so the bot may keep building views there. Start a new chat if you want plain answers back.
If you run a Hermes bot on openclawlaunch.com, open Quick Chat and click Interactive. I wrote up the details and a side by side with ChatGPT here:
My guess is the first thing most people ask for is a calculator. I'd rather hear what you'd make it research.
Top comments (0)