The whiteboard photographer
The architecture diagram is a photo on someone’s phone.
Commit the diagram next to the code.
For engineers · plain files · your own agent
Mumble listens to your design discussion and draws components, calls and trade-offs as you talk. Every arrow cites the sentence behind it. Files are plain JSON; the reasoning runs on your own Claude Code or Codex.
macOS 13+ · whisper.cpp on your Mac · no API key
$ git diff --statdiagram.canvas | 6 +++---Diagrams in git
~/Mumble/2026-09-28-borrow-desk/
capture.json ← the transcript
diagram.canvas ← the map
audio.wav ← never uploaded
- "id": "reminder-queue",
- "label": "Reminder queue",
+ "id": "nightly-reminder-job",
+ "label": "Nightly reminder job",
"source": { "from": "voice", "utterance": "u058" }
Tip: add audio.wav to .gitignore.
The whiteboard photographer
Commit the diagram next to the code.
The “why did we do that?” archaeologist
Every decision box links to the sentence that made it.
The AI-diagram sceptic
Nothing unsaid is drawn as fact. Inference is marked.
Components and labelled calls, drawn with the same ten symbols every time. Switch the style in one click.
See the architecture shape →The options on the table, the trade-offs people named, and the outcome only once someone says it.
See the decision shape →What happened and when, what you learned, and the follow-ups that were actually agreed.
See the retro shape →Point your own Claude Code at ~/Mumble over MCP. The canvas is an API.
Read the MCP docs →$ claude "add a read cache between API and Postgres in borrow-desk" 1 tile added · marked as inferenceMumble never holds an API key.
FAQ
Whichever model your Claude Code or Codex is set to. Mumble doesn’t bring a model of its own for the map; transcription runs locally with whisper.cpp.
The transcript and the canvas, sent through your own Claude Code or Codex account to build the map. Audio never leaves your Mac, and transcription-only mode sends nothing at all.
It’s instructed not to invent infrastructure or draw abandoned options as current. Anything without a spoken source is marked as inference.
Ten: client, API, service, function, database, cache, queue, storage, load balancer and external system. Anything else stays an ordinary box.
Yes. Canvas files are deterministic JSON with readable IDs, so renaming a box is a one-line diff and adding one is a single block. Add audio.wav to .gitignore.
Yes. Point your own Claude Code at ~/Mumble and it can read and edit canvases from the terminal over MCP. It’s a developer setup today.
Not today. The canvas exports as PNG, PDF or to the clipboard, and the summary and transcript as Markdown. The canvas itself is plain JSON if you want to convert it.
Add them to a vocabulary list, for one workspace or one recording, and transcription gets them right.
No. Mumble records your microphone and your Mac’s audio, so it works with Zoom, Meet, Teams or a room with a whiteboard, and nothing joins the call.
Not yet. Canvases are local files, so share them through git for now. Team mode, coming soon, adds sharing and team collaboration.
A Mac on macOS 13 or later, and Claude Code or Codex installed and signed in. Mumble never asks for or stores an API key.
Nothing while Mumble is in beta, and local use on your Mac stays free after it. Team mode, a monthly plan for shared boards, is coming soon.