An open-source meeting assistant: what the source actually lets you check

For a tool that listens to your meetings, open source is not a badge. It is what lets you, or someone you trust, confirm what happens to the audio instead of taking a vendor's word for it. That only works if four things are true.

1. The licence is actually open

“Source available” and “open source” are not the same. An OSI-approved licence means you may read, build, change and redistribute the code. Daisy is Apache 2.0: permissive, with an explicit patent grant.

2. The parts that matter are open

An open SDK around a closed service does not tell you where your audio goes. What needs to be readable is capture, transcription, what is sent over the network and where files are written.

In Daisy all of that is in the app's source: microphone and system-audio capture, on-device transcription through WhisperKit, speaker separation through FluidAudio, the summarizer that talks to the provider you pick, the local MCP server bound to 127.0.0.1, and the list of every host the app can contact. The speech models themselves — Whisper, the pyannote-based diarization models — are open as well.

3. What you run is what was published

Open code does not help if the build you download was made from something else. There are two ways to close that gap: build it yourself, or check the signed build against the source.

Daisy can be built from source in Xcode — clone, open, run; dependencies resolve on first load. Releases are signed with a Developer ID, notarized by Apple and signed for updates. And whatever you run, the ten-minute firewall test on the verify-local page shows what the app does on the network, independently of the source.

4. You can check claims, not just code

Reading a Swift codebase is not realistic for most people. Useful open projects also publish evidence that non-programmers can weigh.

Daisy publishes a reproducible benchmark — a neutral scorer for word error rate and diarization error rate, the exact audio case, and raw output — with its limits stated: one AMI meeting is one case, not a general accuracy claim. Other tools are scored only once their raw output exists for the same audio.

Why it matters for meetings in particular

Meeting audio holds other people's voices, not only yours. When you choose a tool you are choosing for them too. An open, local tool is one of the few ways to make that choice on evidence.