The interesting question is not whether another app can write meeting notes. It is what users gain—and what responsibilities they inherit—when the processing moves onto their own computer.
What Windows users get
Frigade lists Windows 10 version 2004 or newer on x64, Apple Silicon Macs running macOS 14.4+, and Ubuntu 22.04 or similar x64 Linux systems. All require 8 GB of memory. Speaker names and calendar integration are Mac-only.
Features include live transcripts, summaries, 99-language transcription, recording reminders, automatic stopping, folders, search, Markdown export and questions across meetings with source links. Yogurt captures microphone and system audio; call detection supports Teams, Zoom and Meet.
For a Windows evaluation, distinguish basic recording from speaker attribution. Test both sides of a conversation before trusting the app with an important meeting. Bot-free should describe the capture method—not become an excuse to skip notifying participants.
Local AI has a setup cost
The README identifies whisper.cpp for transcription and llama.cpp for summaries and answers. Models vary by hardware. Setup downloads a roughly 5 GB notes model first, then a suitable model potentially exceeding 14 GB; offline processing follows installation.
Plan the trial around that download requirement rather than installing moments before a call. Treat minimum memory as an entry requirement, not a performance promise.
The assistant connection changes the privacy equation
Yogurt’s optional MCP access is disabled by default. Enabling it permits read-only access to meeting records, but there is no per-app token: other software under the same user account can read them. Connected assistants may forward content to online models.
That makes the integration a separate deployment decision, not simply another convenience checkbox. An organization evaluating local processing should first test Yogurt with assistant access disabled, then assess any proposed connection independently.
A sensible first trial
Before adopting Yogurt, use a short, non-sensitive meeting to evaluate:
- Capture: Are both local and remote voices present?
- Accuracy: Do names, numbers and technical terms survive transcription?
- Summary quality: Are decisions distinguished from suggestions?
- Performance: Does processing interfere with the call?
- Portability: Does Markdown export preserve the material you need?
The practical appeal is control over the workflow. The sensible response is neither automatic distrust nor a privacy victory lap: verify the recording, check the notes, and decide deliberately which other applications should be allowed to read them.