Screenshots
About this app
Write !ai what did we decide about the venue? in the room, and a language
model running entirely in your browser answers as a reply, with the recent
conversation as context. No backend, no API key: room content never leaves
your machine. The only network traffic is the one-time download of the model
(from Hugging Face), its GPU code (WebLLM's, from GitHub) and, when first
used, the OCR languages and the voice data (from jsDelivr, at pinned versions).
It summarises the conversation, searches the history, reads voice messages (speech to text), images (what they show and the text in them) and PDFs when asked, and reads answers aloud. Several open widgets in a room share the work: only one answers each request, and the widget shows who keeps a model running for the room.
The model runs on your GPU through WebGPU (WebLLM, in a Web Worker). It needs a browser with WebGPU on a real GPU (current Chrome or Edge); small models are quick but make mistakes. In an encrypted room, the client decrypts messages for the widget and encrypts its answers, so everything stays end-to-end encrypted, and the model sees only what your client already shows you.
Add to a room
Paste this into any room you may configure — Element and most clients understand /addwidget:
/addwidget https://marcade.store/apps/de.datanauten.llm-widget/0.1.2/Needs WebGPU on a real GPU; the default model is a one-time 2.2 GB download.
Rooms should pin the versioned URL above. https://marcade.store/apps/de.datanauten.llm-widget/latest/ always points at the newest release.
Name and icon in Element
/addwidget can carry only the URL: Element lists the app as “Custom” without an icon. Add the Marcade Store once per room and it gives every Marcade app pasted in there its name and icon, and tells you when an update is out.
/addwidget https://marcade.store/apps/de.minagawa.marcade-store/0.1.4/Permissions
Capabilities this widget may request. The Marcade Player asks for your consent before granting any of them.
- Requires running inside a Matrix client
io.element.requires_client - Read messages in the room (type m.text)
org.matrix.msc2762.receive.event:m.room.message#m.text - Read messages in the room (type m.notice)
org.matrix.msc2762.receive.event:m.room.message#m.notice - Send messages to the room (type m.notice)
org.matrix.msc2762.send.event:m.room.message#m.notice - Read messages in the room (type m.audio)
org.matrix.msc2762.receive.event:m.room.message#m.audio - Read messages in the room (type m.image)
org.matrix.msc2762.receive.event:m.room.message#m.image - Read messages in the room (type m.file)
org.matrix.msc2762.receive.event:m.room.message#m.file - Download files and media through the client (e.g. images, avatars)
org.matrix.msc4039.download_file - Send room events of type “de.datanauten.llm_widget.presence”
org.matrix.msc2762.send.event:de.datanauten.llm_widget.presence - Read room events of type “de.datanauten.llm_widget.presence”
org.matrix.msc2762.receive.event:de.datanauten.llm_widget.presence
What's new in 0.1.2
Images: !ai describe says what an image shows (a small vision model in your browser, in English) together with any text in it. Questions about an image now get that description too, instead of OCR text alone.
Information
- Version
- 0.1.2
- Section
- Community — Built by Marcade CI from a pinned commit; the capability list has been audited.
- License
- AGPL-3.0-or-later
- Author
- Yan Minagawa (datanauten)
- Source code
- https://github.com/yncyrydybyl/llm-widget
- Issue tracker
- https://github.com/yncyrydybyl/llm-widget/issues
- Encrypted rooms
- Supported
- Widget ID
de.datanauten.llm-widget