What happened
Frigade’s GitHub repository describes Yogurt as a private AI meeting-notes app and an open-source alternative to Granola. It records microphone input and the sound playing on the computer, without joining the call as a bot. Users can add their own notes while recording. Yogurt offers live transcripts, meeting summaries, search and folders, and answers to questions across meetings with links to sources. It can copy notes or transcripts and export them as Markdown. The project is licensed under MIT and says no account is needed.
The repository lists minimum requirements of macOS 14.4 or later on Apple Silicon, Windows 10 version 2004 or later on a 64-bit PC, or Ubuntu 22.04 or a similar Linux distribution on a 64-bit PC. Each requires at least 8 GB of memory. Transcription uses Whisper through whisper.cpp; supported Macs also use Apple SpeechAnalyzer to show live words, while Whisper supplies the saved transcript. Summaries and answers use a local open-weight model from the Gemma 4, gpt-oss or Qwen3 families, selected for the computer and run through llama.cpp.
Initial setup downloads a smaller notes model of about 5 GB, then the best model the computer can run well, which the README says can be 14 GB or more. Once models are downloaded, transcription and summaries run locally and work offline. Yogurt goes online for model downloads and update checks. The app stores meetings, recordings, settings and models in platform-specific local directories. Multi-language mode supports transcription in 99 languages; English is the default. The repository says summaries follow the meeting’s main language and answers follow the question’s language.
Yogurt can detect calls in Zoom, Teams and Meet and offer to record; it starts recording only after the user chooses. It can record calls in other apps by capturing microphone and system audio. On Mac, it can also read speaker names from Zoom, Teams and Meet with Accessibility permission, and calendar integration is Mac-only. The app says users should tell other people on the call that recording is taking place.
An optional local MCP server lets a chosen assistant app read meetings, notes and transcripts; it cannot change them. Access is off until enabled in Settings. On Mac and Linux, a profile-folder socket uses 0600 permissions. Windows uses a randomly named pipe stored in the user’s profile. The README warns that there is no per-app token: while access is on, any app running under the same user account can read the data. An assistant’s own online model may receive information it reads. Yogurt is free, with no account required.
What the archive adds
New Runtime’s earlier coverage of Draw Things’ Local Code described a Mac app packaging an on-device agent, model downloads and execution tools (earlier coverage). Yogurt is a new event, not a continuation: it applies local models to meeting capture and notes, with an explicit read-only MCP connection to assistant apps.
Trend signal
Yogurt offers a concrete example of software exposing a bounded capability to an external assistant: read-only access to meetings, notes and transcripts through a local MCP server. That supports the direction described in agent-ready software exposes capabilities, while the lack of per-app tokens illustrates a remaining permission boundary. This is one product’s design, not evidence that the approach is broadly established.
Why now
The app combines a familiar meeting-notes workflow with local transcription and summarization, then makes the resulting records available to an assistant through MCP. That matters when teams are evaluating whether meeting data can support agent workflows without sending every recording to a hosted notes service. The practical timing is tied to the release itself; the repository does not state a launch deadline or external event that prompted it.
Why it matters
Yogurt makes the trade-off unusually legible. Teams can keep recordings and generated notes on a computer and continue working offline after model downloads, but local processing does not by itself settle consent, access or endpoint-security questions. Participants still need to know that a call is being recorded, and an assistant may pass material it reads to its own online model.
For builders, the MCP interface is both the useful integration point and the place to inspect carefully. Read-only access limits what an assistant can do to the notes, yet the absence of per-app tokens means access is granted to processes under the same account while enabled. A team assessing the app should test its hardware and language needs, identify which assistants can run under the account, and decide whether their own policies permit recording and subsequent model use. The repository describes capabilities and stated protections, not independent security testing or measured transcription accuracy.
Limits and uncertainty
The repository does not provide measured transcription accuracy, summary quality, latency or resource-use benchmarks across the supported computers. It says model selection depends on hardware but does not specify the precise models or performance users should expect on each device. The README describes local storage and MCP access mechanics; those descriptions have not been independently security-audited here. It also does not quantify how well the app handles noisy calls, overlapping speech or every language among its 99 transcription options. Teams should evaluate the application with representative meetings and their own security and consent requirements.
Skill to build now
Before connecting any meeting-notes app to an assistant, map the data path: identify where recordings and transcripts live, which process can read them, whether access is scoped per application, and whether the assistant sends retrieved text to a hosted model. For Yogurt, test MCP access with a non-sensitive meeting first, keep the integration disabled when it is not needed, and tell meeting participants before recording.
Reader takeaway
Yogurt is a free local-first meeting-notes app whose main appeal is keeping transcription and summaries on the user’s computer after model setup. Its MCP server makes notes useful to assistants, but access is account-wide rather than per-app. Treat local inference and integration permissions as separate questions, and verify both before using it with sensitive conversations.