Frigade launches open-source AI meeting notes for local use

A closed laptop resting on a wooden desk, partially covered by a translucent shield made of frosted glass, with faint sound wave patterns glowing softly beneath it — symbolising local audio processing

Frigade has released Yogurt, a free, open-source desktop application that generates AI-powered meeting notes without sending audio or data to external servers. Designed for Mac, Windows, and Linux, it offers live transcription, summaries, and cross-meeting search — all processed locally. The tool positions itself as a privacy-first alternative to cloud-based services like Granola.

How Yogurt works technically

Yogurt captures microphone and system audio during calls but does not join meetings as a participant. It uses whisper.cpp for speech-to-text transcription, selecting models like large-v3-turbo based on device capability. On supported Macs, Apple SpeechAnalyzer provides real-time word display, while Whisper generates the final transcript. Summaries and Q&A use local open-weight models from Gemma 4, gpt-oss, or Qwen3 families, executed via llama.cpp. Initial model downloads range from 5 GB to over 14 GB, after which full functionality works offline.

Privacy and data control

All recordings, transcripts, notes, and AI processing remain on the user’s device. Data paths are fixed: Mac stores files in ~/Library/Application Support/Yogurt, Windows in %APPDATA%\Yogurt, and Linux in ~/.config/Yogurt or $XDG_CONFIG_HOME/Yogurt. No account is required. The app only goes online for model downloads and update checks. Users must explicitly enable access for external AI apps via a local MCP server, which grants read-only access to meeting content. On Mac and Linux, socket permissions (0600) restrict access to the user’s account; Windows uses randomly named pipes stored locally.

Platform support and limitations

Yogurt supports macOS 14.4+ on Apple Silicon, Windows 10 2004+ (x64), and Ubuntu 22.04+ or similar Linux via AppImage, requiring at least 8 GB RAM. Call detection works across any app by monitoring microphone use and browser tab titles. Speaker identification and calendar integration are currently Mac-only, relying on Accessibility permissions and local calendar access without account sign-in. Multi-language transcription supports 99 spoken languages when enabled, though English is default. The app interface is available in 32 languages.

What it means

For businesses in the UK, UAE, US, and Kazakhstan concerned about data sovereignty and vendor lock-in, Yogurt offers a transparent, auditable alternative to proprietary meeting assistants. Its MIT license allows inspection, modification, and redistribution. However, the large model sizes may challenge users with limited bandwidth or storage. Organisations in regulated sectors — such as finance or legal services — could benefit from avoiding cloud processing, though they must still manage consent and internal policies around local recording.

The reliance on local compute means performance varies by hardware, particularly for larger models. While Windows and Linux are supported, advanced features like speaker labelling remain limited to Mac, reducing cross-platform parity. The MCP integration enables workflow connectivity with local AI tools but requires trust in third-party apps, which can read all meeting data when access is enabled.

Frigade’s approach aligns with growing demand for on-device AI, especially in regions with strict data laws. Unlike services that require subscriptions or cloud accounts, Yogurt’s model supports user control and long-term accessibility without dependency on a vendor’s uptime or pricing changes.

Sources

Leave a Reply

Your email address will not be published. Required fields are marked *