Gradio 6 quietly shipped a very powerful feature: gr.HTML now supports custom templates, scoped CSS, and JavaScript interactivity. Which means you can build pretty much any web component — and Claude (or any other frontier LLM) can generate the whole thing in one shot: frontend, backend, and state management, all in a single Python file.
We tested this by building different types of apps. Each one is a single Python file, no build step, deployable to Hugging Face Spaces in seconds.
\nPomodoro Timer: A focus timer where a pixel-art tree grows as you work. Starts as a seed, sprouts branches, grows leaves. Complete a session and the tree joins your forest. Session tracking, theme switching, break modes — all interactive, all in one file.
\nThe tree animation alone would normally require a separate React component. Here it's just CSS keyframes in css_template and state updates in js_on_load.
GitHub Contribution Heatmap: Click any cell to toggle contributions. Multiple color themes. Pattern generators (streaks, seasonal, random). Live stats that update as you edit.
\nKanban Board: Full drag-and-drop between columns. Inline editing (double-click any card). Search feature that can filter in real-time. Collapsible columns.
\nDrag-and-drop usually means pulling in a library. Here it's native HTML5 drag events wired up in js_on_load, with state synced back to Python via trigger('change').
Spin-to-Win Wheel: Smooth CSS animation, rotation state that persists across re-renders. Preset configurations for yes/no decisions, restaurant picking, team selection. You can also add custom spinning segments on the fly.
\nThis is where gr.HTML gets really interesting for ML work: you can build specialized components that can handle your exact output format, then use them like any built-in Gradio component.
Detection Viewer: A custom viewer for object detection, instance segmentation, and pose estimation results. Renders bounding boxes, segmentation masks, keypoints, and skeleton connections — all in a reusable gr.HTML subclass that plugs directly into your model pipeline.
The community's built some creative components with gr.HTML too:
\n3D Camera Control for Image Editing: A full Three.js viewport inside a Gradio app! Drag handles to control azimuth, elevation, and distance. Your uploaded image appears in the 3D scene, and the camera parameters feed directly into Qwen's latest image editing model. These kinds of interactive 3D controls would normally require a separate frontend — with Gradio it's just one gr.HTML subclass🔥
Real-time Speech Transcription: Live transcription with Mistral's Voxtral model. The transcript display is a custom gr.HTML component with animated status badges, a live WPM counter, and styled output that updates as you speak. Real-time UI feedback without using React!
Every gr.HTML component takes three templates:
\n${value} injects Python state. props.value updates it from JavaScript. trigger('change') syncs back to Python. That's the whole API.
For reusable components, subclass gr.HTML:
\n\n \nNow Heatmap() works like gr.Image() or gr.Slider() — use it in Blocks, wire up event handlers, whatever you need.
When you ask your LLM to build a custom component, single-file output is everything. No "now create the styles file" or "wire this into your build config." Just one Python file that runs immediately.
\nThe feedback loop becomes: describe what you want → get code → gradio app.py → see it working → describe what to fix → repeat. Each cycle takes seconds with Gradio's reload mode.
Deploy to Spaces with gradio deploy or share a temporary link with demo.launch(share=True). Within a few seconds from an idea to a live app.
Gradio ships with 32 interactive components, but sometimes your perfect AI web app needs something special. That's where gr.HTML comes in.
If you’ve got an idea, try building it with gr.HTML: describe what you want to your LLM, generate the code, run it. You might be surprised what you can ship in 5 minutes.
Suggested reading:
\n