FrigadeHQ/yogurt

Yogurt, built by Frigade, records microphone and system audio, transcribes it in real time using whisper.cpp (plus Apple SpeechAnalyzer on supported Macs), and then uses llama.cpp to run locally selected open-weight models such as Gemma 4, gpt-oss, or Qwen3 to generate summaries, answer cross-meeting questions with citations, and more. It identifies speakers in Zoom, Teams, and Meet, supports transcription in 99 languages in multilingual mode, and allows exporting notes and transcripts to Markdown. Once models are downloaded, they can be used offline.

Yogurt, built by Frigade, records microphone and system audio, transcribes it in real time using whisper.cpp (plus Apple SpeechAnalyzer on supported Macs), and then uses llama.cpp to run locally selected open-weight models such as Gemma 4, gpt-oss, or Qwen3 to generate summaries, answer cross-meeting questions with citations, and more. It identifies speakers in Zoom, Teams, and Meet, supports transcription in 99 languages in multilingual mode, and allows exporting notes and transcripts to Markdown. Once models are downloaded, they can be used offline.

Why it matters A free, open-source alternative to Granola, Otter, and Fireflies that requires no account registration; it can expose a local MCP service, allowing coding agents to read your meeting notes in read-only mode.

Stars 2778 (+153 / 24h)

Reasons

  • 153 new stars in 24 hours, total stars 2778 (indicating attention only); MIT license
  • The README specifies the data storage directory, notes that MCP access is disabled by default and read-only, and details permission settings for the local socket

Usage

  • Download the installer for Mac (Apple Silicon), Windows, or Linux from Releases; on Mac, you can also run brew install --cask frigadehq/yogurt/yogurt. On first launch, it downloads transcription and note-taking models based on your computer's configuration, requiring at least 8 GB of RAM

Risks

  • Recordings are saved by default, and there is no bot-join notification, so you must inform other participants before recording
  • Once MCP access is enabled, any application running under your user account on the local machine can read meetings, transcripts, and participant names and email addresses; connected assistants may also send content to their own online models

Open

GitHub

Explore the latest AI Radar updates