Put your MacBook on the table, or let Huddle listen to the other side of a call. Press stop and
the notes write themselves — while nothing leaves your computer.
The team reviewed the install experience. Bundling the speech models pushes the download
past 5 GB, so they will ship the app alone and fetch the recommended models on first run,
with the onboarding rewritten to explain the wait up front.
Decisions
Models download on first run instead of shipping inside the app bundle.11:52
Keep the menu-bar recorder in the release; it tested well.23:14
Action items
Find action items
Measure the bundle size without models Dylan09:41
Rewrite the onboarding copy for the model download
Maya Friday12:04
Transcript
Click a speaker to rename · click text to seek
Maya11:38
If the models ship inside the app, the first download is enormous. Can we pull them after install instead?
Dylan11:52
Yes — we fetch the recommended set on first run, and let people point Huddle at whatever they already have in Ollama.
Speaker 312:04
Then I'll rewrite the onboarding before Friday so it explains the download up front.
Why Huddle
Every other notetaker uploads the room.
Meeting AI usually means a bot joining your call and your conversation living on
someone else's server. Huddle does the whole job — transcription, speaker separation,
summaries — on the machine already sitting in front of you. Turn off the Wi-Fi once the
models are downloaded and it keeps working.
Audio recordingstays on disk
Transcript & speakersstays on disk
Summaries & action itemsstays on disk
Sent to a servernothing
Account requirednone
What it does
Record the room. Get notes you'd actually send.
Recording
In the room or on a call. Captures the microphone and, optionally, the audio of every other app — Teams, Zoom, Meet, FaceTime.
Menu bar mode. Close the window and a small recorder stays in the menu bar; ⌥⌘R starts and stops from anywhere.
Crash-safe. The file is finalised every second, so a dead battery still leaves a recoverable recording.
Import. Drop in existing audio and it gets the same treatment.
Notes
Who said what, with timestamps. Rename or merge speakers; Huddle suggests names it recognises from earlier meetings.
Summary, topics, decisions with evidence links back into the transcript — in any of 56 languages, whatever was spoken.
Action items extracted on demand and tracked across meetings until you tick them off.
Refine and Ask. Tell Huddle what it got wrong and it rewrites; ask one meeting, or all of them, a question.
Private by design
Everything runs locally — transcription, speaker separation and summaries. Fully offline once the models are downloaded.
Your data stays yours, in one folder under Application Support. Export any meeting as Markdown, text, SRT, JSON or the original audio.
Storage limit. Set how much disk recordings may use; the oldest audio is pruned first.
Works with your AI tools
MCP server. Claude Desktop, Claude Code, Cursor and other MCP clients can search your meetings, read transcripts and pull open action items — without touching your filesystem.
Optional network access. Share the MCP server on your local network with API keys.
Bring your own models. Huddle finds what you already have in Ollama, LM Studio or the Hugging Face cache and downloads only what is missing.
For the terminal crowd
Your meetings, in your assistant.
Huddle ships an MCP server over the same services the app uses. Point Claude Code, Claude
Desktop or Cursor at it and ask about last week's call without handing anyone a file.
Settings → MCP prints the ready-made configuration for each client.