Autorec

Every meeting recorder on the market solves this by sending a bot into your call and your audio into their cloud. If you handle client work, health data or anything legal, that ends the conversation. A privacy claim only means something if the recording never leaves the machine, which puts transcription, summarisation and encoding on the user's own laptop, in real time, without spinning the fans up.

The dependency tree mattered as much as the code. One GPL library anywhere in it and the product cannot be sold.

Native C++20 with a webview frontend, so the capture path is real code and the UI is still quick to build. Recording runs on four threads with a zero-allocation audio path; frames go BGRA to YUV420P to H.264/AAC into MP4 without any copy that wasn't needed. Meeting detection watches for the conferencing apps themselves, so nothing ever has to join the call.

Transcription is whisper.cpp, on-device, with subtitle output. Summaries go through any OpenAI-compatible endpoint with a prompt the user can edit. The one feature that genuinely wants a network is also the one you can repoint or switch off.

Dependencies got picked for their licence before their capability: screen_capture_lite, miniaudio, LGPL FFmpeg. Licensing is one-time keys, five devices each, re-checked weekly and valid offline. An app that stops working on a plane isn't a local-first app.

Shipping on Linux and Windows, packaged five ways out of CI, with a macOS port underway. It records, transcribes and summarises with the network unplugged. The licence audit that makes it sellable was settled before the first line of capture code got written.