In my company I am an engineering lead, which at least for me typically means 4-5 or more meetings a day. That's fine and part of the job, but at least for me it's sometimes hard to keep track of things, what was decided, what was promised, what other people told me and what I told them. Most tools I tried are either insufficient (pure transcription and recording) or put the audio and data on someone elses computer. I wanted neither, so a while ago I started building the thing I wanted.

Working from home, most of my todos either stem from Slack or from meetings, and that was how it started: Record the meeting on app & microphone, and transcribe them with open models (in my case Whisper Large v3 Turbo, diarization with pyannote). This works reasonably well, and is not something new in itself. I also copy my slack threads into the system.

Now, the actual use of it comes after. You confirm who was in the meeting and who spoke what (it integrates with your Apple Calendar) and clean up after the transcription and diarization. The result is then heavily used downstream, an LLM (local, mileage can vary depending on your machine, or cloud) will extract todos, topics discussed, who said what and who promised what to whom. Because we cleaned up first, you then get an inbox full of things - you can look up what person X actually promised you, what facts about them were extracted, what topics are discussed during your conversations. All of those are auto extracted, but can be confirmed or rejected by you after parsing the meeting.

The app also produces everything in Markdown in a folder of your choice. This means what it produces can be processed by anything else in your workflow as well, and so if you choose to stop using horch, nothing is really lost.

It of course also has other interfaces - it has a CLI to integrate with, an MCP to connect against, so that it is as open to your other workflows as possible. This way for example, you can take the context of a meeting, throw it against a codebase or another source of information, and connect the two.

Right now, it's Macs only. It is paid (49€ once, with a year of updates), but has a trial phase of 30d if you wanna try it with no details needed, and the version you have will work forever. I did not open source it. I am not financially dependent on it and I will keep building it for myself either way, but I think it can be worth paying for, and if someone wants to support the work, I would be very grateful.

I would love for you to try it, and I am most interested in where it breaks. What are your pain points, what would you like to see.