Skip to content
← Back to projects

Whiteboard

A real-time collaborative whiteboard with an AI Assistant that draws on the Board with you: editable Mermaid diagrams, clickable app Prototypes, voice requests, and Live calls it maps into notes and diagrams as people talk. Self-hosted, on any OpenAI-compatible model.

Created · Updated
typescriptreactbunexcalidrawyjsaiself-hosted

A shared canvas for a team, with an Assistant that works on the same Board you do. Every change it makes is ordinary Board content, so people can move, edit, undo or redirect it like anything else.

What it does

  • Draw together. Excalidraw kept in sync over Yjs: presence, follow and “follow me”, sticky notes, Team chat, and Owner, Editor and Viewer roles enforced on the server for every change.
  • An Assistant on the Board. Ask in your own Conversation and it reads the Board, then adds shapes, sticky notes, templates and Mermaid diagrams that arrive as editable shapes. Each change is listed in words with Undo and Redirect.
  • Clickable Prototypes. Describe an app and the Assistant builds Screens of real UI components with working navigation, forms and sample data. Edit them by hand with layers and an inspector, see every Screen at Desktop, Tablet and Mobile at once, and click through in Play while others follow along.
  • Talk to it. Voice requests are transcribed and the reply is read back a sentence at a time, interruptible mid-sentence.
  • Live calls, mapped. Editors talk on a call; every utterance is captioned by name, and Live mapping digests the conversation every couple of minutes into running meeting notes and diagrams, each node citing who said it. It can also move tickets on a linked sprint board during a standup.
  • Bring your Excalidraw files. Import a .excalidraw file as a new Board, or drop one onto a Board to add it beside what’s there.

How it’s built

Bun serves a React app and the Board’s Yjs sockets from one binary, with SQLite for everything stored. Every incoming change passes one enforcement module, so roles, Prototype specs, images and links are checked whoever sends them, the Assistant included. The Assistant runs on any OpenAI-compatible endpoint; mine is a local model on a Mac Studio, with speech-to-text and text-to-speech on the same machine, rather than a hosted API. Board tools run in a worker, and every run is undoable change by change.