FrigadeHQ/yogurt

Yogurt records microphone and system audio without adding a bot to the meeting: it uses whisper.cpp (plus Apple SpeechAnalyzer on supported Macs) for real-time transcription, and llama.cpp to run local open-source models for generating meeting minutes, answering questions across meetings, and providing citations. Multilingual mode supports transcription in 99 languages; on Mac, it can extract speaker names from Zoom, Teams, and Meet; minutes and transcripts can be exported to Markdown. When enabled in settings, a local MCP service allows assistant apps to read-only access meetings, notes, and transcripts.

Yogurt records microphone and system audio without adding a bot to the meeting: it uses whisper.cpp (plus Apple SpeechAnalyzer on supported Macs) for real-time transcription, and llama.cpp to run local open-source models for generating meeting minutes, answering questions across meetings, and providing citations. Multilingual mode supports transcription in 99 languages; on Mac, it can extract speaker names from Zoom, Teams, and Meet; minutes and transcripts can be exported to Markdown. When enabled in settings, a local MCP service allows assistant apps to read-only access meetings, notes, and transcripts.

Why it matters Meeting recordings and minutes do not need to be uploaded to the cloud, work offline without an account, and allow your AI assistant to read-only access meeting records via local MCP.

Stars 958 (+476 / 24h)

Reasons

  • 476 new stars in 24 hours (indicating attention only); the repository was created on 10 11, so reading the README content is normal
  • The README's FAQ clearly specifies data storage locations, model download sizes, and permission boundaries for MCP access

Usage

  • Download the installer for Mac, Windows, or Linux from the official website; on first run, models are downloaded based on your computer's configuration, after which transcription and minutes can be used offline

Risks

  • Requires at least 8 GB of RAM; the initial minutes model download is approximately 5 GB, while better models may exceed 14 GB
  • Recording does not notify other participants, so you must inform them yourself; while MCP access is enabled, any application under the same account can read the data, and connected assistants may send content to their own online models

Open

GitHub

Explore the latest AI Radar updates