RXed AI News

The private meeting-notes machine: record, transcribe, summarise — on your own laptop

Minutes write themselves, and the recording never leaves your machine

A meeting recorder, transcriber and minute-writer that runs entirely on the computer you already own — because your client conversations are nobody else's training data.

The problem

Every client meeting produces a second, invisible job: typing up what was said, what was decided, and who does what. Done properly it costs 20–30 minutes per meeting; skipped, it costs a forgotten commitment three weeks later. The popular fix — cloud note-takers like Otter, Fireflies or Fathom — works, but at two prices. The visible one: roughly €8–20 per user per month, forever. The invisible one: a bot listens to your client conversations and stores the audio and transcript on someone else's servers. For a consultant, accountant or agency handling confidential client matter, that second price is the real problem — and it's exactly the one a local setup eliminates. In 2026 this is a solved problem: on-device speech recognition is fast and accurate enough that the whole pipeline — record, transcribe, summarise — runs on an ordinary office laptop.

The architecture

One computer, two programs. Meetily owns capture and transcription; Ollama owns everything language — summaries, action-point extraction, questions against old transcripts. Both are free, open source, and offline. There is no server to rent for this blueprint: the machine you take to meetings is the infrastructure.

Tool choices — and why

Meetily (MIT, 22.1k stars, v0.4.0 released 05/06/2026) is the pick because it's the rare local-AI tool packaged for normal people: a signed installer for macOS and Windows, ten minutes from download to first transcript, no Docker, no terminal. It captures system audio directly — so no awkward "notetaker bot" joins your client's call — transcribes live with Whisper (or the faster Parakeet engine), and hands the transcript to any Ollama model for summaries. The honest caveats: the Community Edition is young software (v0.4.x, an active issue tracker to match), Linux means building from source, and speaker labels are best-effort — read them as "probably who said what", not court transcript. There's a paid Pro tier (~$10/user/month); the free MIT edition does everything this blueprint needs.

whisper.cpp (MIT, 50.3k stars, active — last push 31/05/2026) is the transcription engine inside. It matters because it's the most battle-tested on-device speech recognition there is, runs well on Apple Silicon and ordinary CPUs, and handles Dutch, French, German and ~90 other languages with the multilingual models — relevant the moment your meetings aren't in English.

Ollama (MIT, ~175k stars, active) runs the summary model locally. A small instruct model (3–8B) is enough for minutes and action points on a 16 GB machine. If you already run Ollama for another blueprint in this library — the SOP answer machine, the invoice chaser — it's the same install doing one more job.

Monthly cost. Software: €0, all three MIT-licensed. Hardware: the computer you already have — an Apple Silicon Mac or a reasonably modern Windows machine; 8 GB RAM works for transcription alone, 16 GB is the comfortable floor once Ollama writes the summaries. The cloud alternative at, say, 3 people is roughly €300–700/year, every year — plus your client audio on a third party's servers. Here the running cost is electricity.

Setup outline

1. Download the Meetily installer for your OS from the GitHub releases page and run it.
2. Install Ollama and pull one small instruct model suited to your RAM (a 3–4B model on 8–16 GB; a 7–8B model if you have headroom).
3. In Meetily's settings, point the summary provider at your local Ollama and pick the model.
4. Choose the Whisper model: if any of your meetings are in Dutch, French or another non-English language, pick a multilingual model (small or medium), not the English-only default.
5. Set the notes storage folder to somewhere your normal backup already covers.
6. Dry-run: record two minutes of yourself stating three fake decisions, and check the transcript and the summary catch all three.
7. Write your one consent line and use it at the start of every recorded meeting: "I'm recording this for the minutes — it stays on my machine; want a copy afterwards?"
8. Pilot for two weeks on internal meetings before you rely on it for client work.

Pitfalls — the real ones

Recording consent is law, not politeness. In the EU you must inform participants (GDPR applies to voice); several US states require all-party consent. The one-sentence announcement in step 7 costs nothing and keeps you clean. Local storage helps your case — but it does not replace asking.
The microphone decides the quality. A laptop mic across a meeting-room table produces a transcript full of holes, and the summary inherits every hole. A €30–60 USB conference mic is the single highest-leverage purchase in this whole blueprint.
Summaries are drafts. Local models occasionally mangle names, numbers and dates — exactly the things that matter in minutes. Proofread every name, amount and deadline before minutes go to a client. The transcript is your ground truth; keep it.
Wrong Whisper model, silent failure. The English-only models will happily "transcribe" a Dutch meeting into English-flavoured nonsense. If meetings are multilingual, this is the first setting to check, not the last.
Keep the audio until the minutes are approved. Young software plus important meeting is a bad place to discover a truncated transcript. Delete recordings on a schedule you choose — after the minutes are confirmed, not before.
Don't let the archive become the leak. The point is confidentiality: full-disk encryption on (FileVault / BitLocker), and the notes folder excluded from any cloud-sync tool you run.

Verified repos

Meetily — MIT, 22.1k stars, v0.4.0 (June 2026)
whisper.cpp — MIT, 50.3k stars, active (May 2026)
Ollama — MIT, ~175k stars, active (June 2026)