Cloud notetakers are great until you think about where your meeting audio actually goes. Granola and the rest are slick, but every word of your call gets shipped to someone else's servers, run through someone else's model, and billed to you monthly. I wanted the opposite trade. So I built Kōchi.

Kōchi listens to your meetings, transcribes both sides of the conversation, tracks your goals, and coaches you in real time. And it does all of it on your Mac. No cloud. No account. No API keys. No subscription. Your audio never leaves the device.

_{Live coaching and goal tracking · saved meeting detail · search across every transcript.}

Watch it work

Two minutes, start to finish. Set your goals, hit start, and watch the transcript fill in both sides of the call while the coach nudges you along.

What it actually does

It captures both sides of a call. Your microphone shows up as Me: and the far side — system audio, pulled in through ScreenCaptureKit — shows up as Them:. The transcript reads like a real conversation instead of a one-sided wall of text.

It tracks up to three goals per meeting. Type something like "Confirm the launch date" and Apple's on-device model marks it hit the moment the conversation covers it. Don't agree? Tap the goal and override it yourself.

It coaches you live, with a coach video and prompts that react to how the meeting is going. When you hit End, the whole session saves locally — transcript, goals, and a mixed mic-plus-system-audio recording. Then you can search across every meeting you've ever had and jump straight to the moment you need.

The interesting part

The coaching and goal analysis run on Apple's new Foundation Models framework. One import FoundationModels and you get a real language model running on the Neural Engine, no server in the loop. The transcription is Apple's Speech framework. The audio mixing is AVFoundation. The whole thing is SwiftUI, native to macOS, with zero third-party dependencies. Apple frameworks only.

That's the bet: the local model is finally good enough to do the work a cloud LLM was doing a year ago. No round trip, no rate limit, no bill. It works on a plane.

The catch

Foundation Models ships only in macOS 27, which is still in developer beta right now. So Kōchi needs an Apple-silicon Mac with Apple Intelligence turned on, the macOS 27 beta, and an Xcode with the matching SDK. Early, but that's where on-device AI lives, and this is the first thing I've built that genuinely needs it.

It lives in the menu bar — click the icon and the window comes to the front, reopening one if you've closed everything. Free, open source, MIT. The code is on GitHub.

The whole industry is busy moving your conversations to the cloud. Kōchi keeps them on your desk, where they belong.