Why Meeting Transcription Should Stay on Your Mac

The first time I joined a meeting and saw a little bot pop into the call to “take notes for me,” I sat with the feeling for a while. I had agreed to that. Somewhere in onboarding, I had clicked through a permission. But nobody else on the call had. The bot was listening to all of us and sending all of us into a system I did not own.

That moment is the entire reason lognote exists.

What is the actual default?

Pick any meeting tool that does AI summaries today. Almost without exception, the path your audio takes is this: your microphone, your laptop, an upload to their cloud, a transcription model running on their hardware, a summary stored in their database, and eventually a payload that gets sent back to you. The recording sits on a server you do not control, governed by a privacy policy you skimmed once.

That is the default. Not a worst case. The default.

We have collectively normalized this because the alternative felt impossible. Transcription was hard. Summaries were harder. Running anything serious on a laptop was a non-starter five years ago. So we shipped our conversations to other people’s computers and called it modern.

The hardware caught up. The software didn’t.

What does local-first actually mean?

The phrase gets stretched into shapes it shouldn’t take, so here’s the precise version.

Local-first, for lognote, means:

Your audio never leaves your Mac. Transcription runs on-device with MLX-Whisper.
The recording and the transcript live in your filesystem, as plain Markdown in your vault.
If you choose a cloud model to write the summary, that is your choice and your API key. The tool brokers the call, but the relationship is yours.

What it does not mean: anti-cloud purism. I use cloud services every day. I have nothing against the people building them. The argument is not that the cloud is bad. The argument is that the default should be different for this specific kind of data.

Why this kind of data?

A meeting transcript is not a search query. It is not a photo of your lunch. It is, in most cases, the closest a digital tool will ever come to overhearing you think out loud with another human.

It contains:

Half-formed ideas you would never publish.
Quotes from people who did not sign up to be quoted.
Things you said about your boss, your competitors, your team.
Numbers, plans, and timelines that haven’t been announced.

This is not “user data.” This is conversation. The question of who gets to hold a copy of your conversations should not have a default answer that benefits everyone except the people doing the talking.

Here is the part I keep coming back to. When you turn on a cloud meeting recorder, you are not just consenting on your own behalf. You are consenting on behalf of every other person in the room. The host clicks one button and four other people are now in someone else’s database.

Most of them will never know. The ones who do know have no realistic way to opt out without making the meeting awkward.

Recording laws vary, and how you handle consent in a meeting is on you. What software can change is the math underneath. If the recording stays on my Mac, no third party heard the people I was talking to. That is a smaller promise than “your data is safe,” and a more honest one.

The trade-off

Local-first is not free. A cloud transcription service can run on a beefy GPU cluster and stream you results in real time. A laptop has to do the same work with the chip it has. There are meetings where the on-device transcription takes a few minutes to finish after you hang up. There are summaries that won’t be as polished as the one from a frontier model unless you bring your own key to one.

That is the trade-off. I’m okay with it. You might not be, and the honest version of that conversation is the one I’d rather have than a marketing one.

The reframe

The thing I want from a meeting tool is not “best summary.” It is “summary I am willing to take responsibility for.” If something the model wrote ends up in a document, a Slack thread, or a client email, I’m the one who put it there. The model is a tool. The transcript is a record. The choice of where that record lives is mine.

Local-first is just the version of that choice where the answer doesn’t require me to trust anyone I haven’t met.

Next time you finish a meeting you were allowed to record, open the resulting transcript in your notes folder and ask one question: would I be comfortable if a stranger at another company had a copy of this? If the answer is no, you already understand why a tool like lognote needs to exist.