expectedwrong hindsight

Someone Built My Thing, Except It Works

The right architecture for meeting AI has been obvious for a while — grab system audio and don't ask permission from Zoom.

2 min read 228 words #ai #tools #audio #meetings #mac
hindsight — nailed it

System audio capture remained the correct approach. The bot-joins-the-meeting model stayed fragile. Someone else built the right version because the right version was always simpler.

There's a Mac app making the rounds that records your meetings by hijacking the audio — not by joining as a bot, not by asking the platform's permission, not by praying the Zoom API doesn't break again. It just grabs whatever's coming out of your speakers at the OS level and runs from there.

This is the correct approach. I know this because I built the wrong approach first.

The bot-joins-the-meeting model is fragile in a way that only becomes obvious after you've spent three hours debugging why your Google Meet integration stopped working on a Tuesday for no reason. Every platform is a different API surface, a different authentication flow, a different set of undocumented rate limits. You are permanently one product update away from being broken.

System audio capture doesn't care what meeting software you're running. Zoom, Teams, Google Meet, some obscure WebRTC thing your client insists on using — irrelevant. The audio comes out of the speakers. You take it. Done.

Mac-only is a real limitation and also kind of not, given who actually uses these tools.

What I find interesting is that the "obvious" architecture — the one that sidesteps the entire problem — kept getting passed over in favor of the bot approach, probably because it feels less legitimate. Like you're sneaking. But you're not sneaking. You're just listening to your own computer.