Yogurt keeps AI meeting notes on your own machine
FrigadeHQ's MIT-licensed desktop app records calls, transcribes them and summarizes them with models running locally, and it gathered 1,328 GitHub stars in three days.
A local AI meeting notes app that skips the bot
FrigadeHQ/yogurt, a local AI meeting notes app that keeps recordings and AI processing on the user's computer, reached GitHub trending with 1,328 stars and 62 forks as of 2026-11-14. The repository was created on 2026-11-10, so it collected that attention in three days. It is written in TypeScript, licensed MIT, and pitched as a free, open-source alternative to Granola for Mac, Windows and Linux. No account is needed. The project's repository documents the details.
What the app does
Yogurt captures your microphone and the sound your computer plays, so no bot joins the call. From there it produces a live transcript and a written summary that folds in any notes you type during the meeting. Recordings, transcripts and notes are stored locally.
The smaller features matter as much as the headline ones:
- A recording prompt when it detects a supported call, which you can decline.
- Automatic stop when the call ends.
- Speaker names read from Zoom, Teams and Meet, on Mac only.
- Multi-language transcription covering 99 languages, with English as the default. The app interface is offered in 32 languages.
- Folders and search across meetings by title, notes or transcript.
- Questions answered across meetings, with links back to the source meeting.
- Copy, or export notes and transcripts as Markdown.
How the AI stays on the device
Transcription runs through whisper.cpp. On supported Macs, Apple SpeechAnalyzer supplies live words while Whisper produces the transcript that is saved. Summaries and answers come from a local open-weight model drawn from the Gemma 4, gpt-oss or Qwen3 families, chosen for the hardware and run through llama.cpp.
Setup is the one heavy moment. Yogurt first downloads a smaller notes model of about 5 GB, then the best model the machine can run, which the project says can be 14 GB or more. After that, recording works and notes and questions run offline. The app goes online for model downloads and update checks only. Meeting data lives in ~/Library/Application Support/Yogurt on Mac, %APPDATA%\Yogurt on Windows, and ~/.config/Yogurt (or $XDG_CONFIG_HOME/Yogurt) on Linux.
The hardware floor: macOS 14.4 or newer on Apple silicon, Windows 10 version 2004 or newer on x64, or Ubuntu 22.04 or newer on x64, each with at least 8 GB of memory. Setup tells you if the computer is too slow.
What it means for people building apps
Two design choices are worth studying. First, pushing inference to the client removes per-user model cost and the privacy question at once, trading both for disk space, memory and a long first-run download. Second, a local MCP server lets other assistant apps read meetings, notes and transcripts, and it is off until you enable it under Settings then AI apps.
The permission model is documented with unusual frankness. Mac and Linux use a socket in your profile folder with 0600 permissions, and Windows uses a named pipe with a new random name on each launch. There is no per-app token, so any app running under your account can read while access is switched on, and connected apps appear under names they report themselves. That is a defensible trade for a personal desktop tool, but it is not a sandbox.
Limits and open questions
Speaker names and calendar access are Mac only for now, and the project gives no date for the other platforms. The Windows build is not signed with a Microsoft certificate, so SmartScreen shows a warning until you click through. The source offers no accuracy figures for transcription or summaries, no measured memory use per model tier and no release cadence beyond the MIT license. Star counts on a three-day-old repository measure attention, not reliability.
The builder takeaway
If you are weighing local models against hosted APIs inside your own product, the constraints Yogurt documents, from download size to memory floor to capability per model tier, are the same ones you will run into. Draftship's guide to adding AI to your app covers where a model call fits into an app's logic.
Primary sources
- FrigadeHQ/yogurt: Private AI meeting notes that stay on your computer. A free, open-source alternative to Granola for Mac, Windows and Linux. · GitHub trending AI repositories, Nov 14, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.