Do AI meeting notes work offline on a Mac?
Recording and transcribing with the wifi off is a solved problem on a Mac. The summary is the part that breaks, because a summary needs a language model and most meeting-notes apps keep theirs in a datacentre.
- Offline transcription is no longer unusual — MacWhisper, Superwhisper and Narada all run the speech model on your Mac. Offline summarising is the rare part, because it needs a language model on the same machine.
- The cloud notetakers cannot do it at all by design. Granola’s own security page explains why: “we initially tried doing these locally on device, but the computation was too much and it slowed down your computer.”
- Otter’s offline capture is a phone feature — the mobile app “can capture audio locally on your device when you’re offline” and uploads later. On the Mac it documents none, and transcription “requires a stable internet connection”.
- Narada’s On-device AI mode runs the summary on your Mac too, with real gates: Apple Silicon only, a one-time 3 GB or 7 GB model download, 8 GB RAM for the smaller model and 16 GB for the larger, and lighter notes than the cloud route.
- Offline does not mean feature-complete. Calendar sync needs the network, the models need one download, and on-device analysis pauses while a recording is running.
Recording a meeting and turning it into text with the network off is a solved problem on a Mac. Several apps do it, and they do it well. What almost nothing does offline is the next step — reading the transcript and writing the summary — because that needs a language model sitting on your own disk rather than in someone’s datacentre.
So “does it work offline” is really three questions with three different answers.
The three stages, and which one needs a network
Capture never needs a network. Your Mac is already playing the other person’s voice and already hearing yours; writing both to disk is local by definition. Even the cloud services record locally first.
Transcription used to need a network and now does not. Apple Silicon runs a speech model fast enough that uploading the audio buys you nothing, and a handful of Mac apps have moved it on-device. Narada uses Parakeet TDT v3 through Core ML — a one-time download of roughly 600 MB, after which no audio leaves the machine in any mode.
Summarising is where the line falls. A summary is a language-model job, and a useful model is gigabytes of weights. Vendors who process in the cloud decided that trade explicitly. Granola, to its credit, publishes the reasoning on its own security page:
“For transcription and AI summarization, we process this in the cloud to provide the best quality - we initially tried doing these locally on device, but the computation was too much and it slowed down your computer.”
That is a real engineering position, not a dodge. It is also a decision worth disagreeing with in 2026, on a laptop with 32 GB of unified memory.
The airplane-mode table
What each app can actually do with wifi off, from the vendors' own documentation.
| App | Records | Transcribes | Summarises | Where the summary model lives |
|---|---|---|---|---|
| Narada, On-device AI | Yes | Yes | Yes, on Apple Silicon | Downloaded once, managed by the app |
| MacWhisper | Yes | Yes — “all performed locally on your Mac” | Only via your own local model | Ollama or LM Studio, which you install and run |
| Superwhisper | Yes, meeting mode | Yes, with local voice models — a Pro feature | No — every language model it lists is cloud | Superwhisper, Anthropic, OpenAI or Groq, all Pro |
| Meetily | Yes | Yes, Whisper.cpp | Yes, with local Ollama models | Self-hosted; needs Node, Python and FFmpeg |
| Granola | Capture is local | No — live socket to Deepgram or AssemblyAI | No | Cloud, by stated design |
| Otter | On mobile only | No — after reconnecting | No | Cloud |
Two rows deserve a sentence each.
Otter's offline story is a delay, and it belongs to the phone. The one offline path in its documentation is the mobile app, which “can capture audio locally on your device when you’re offline. After recording, reopen the app when you’re back on Wi-Fi or a strong cellular connection so the upload can finish.” Nothing equivalent is documented for the Mac desktop app, and transcription “requires a stable internet connection … a minimum uplink bandwidth of 512kbps”. The transcript exists once you have signal again. That is a reasonable design; it just is not a tool for a plane, and on a Mac it is not a documented one at all.
Granola has nothing to retry from. Its transcription runs as a live WebSocket to api.deepgram.com and streaming.assemblyai.com, and its docs also state that it “does not record or save audio or video at any point during the call, so there’s no way to access audio from your meetings.” No audio file means no queued upload the way Otter has one. Granola documents caching notes locally for offline reading — that is a different thing from producing them.
“Offline summaries” splits into two very different products
Three options in that table can produce a summary with the network off — Narada, MacWhisper and Meetily. They ask you for wildly different amounts of work first.
The bring-your-own-model route is real and it is cheap. MacWhisper’s privacy documentation states the exception plainly: “if you use a local AI Service Provider such as Ollama or LM Studio, which runs locally on your Mac. That way, MacWhisper can get AI Prompt responses without using the internet.” Meetily goes further and is MIT-licensed, but its own minimum requirements are “8GB RAM, 4-core CPU, 10GB storage; Node.js 18+, Python 3.10+, FFmpeg” — a stack you maintain.
Narada ships the model instead. Turn on On-device AI and it picks a checkpoint that fits your machine, downloads it once, and runs it in-process on Metal. No second app, no local server, no port, no API key, no account. The choice is made by hardware, not by you: 8 GB of RAM gets Gemma 3 4B at about 3 GB on disk, 16 GB or more gets Gemma 3 12B at about 7 GB.
The whole intelligence layer follows the same rule. Search and chat over the archive use a small on-device embedding model — about 120 MB — so asking “what did we agree about the migration” works with the network off too, and the answer still cites the meeting, the speaker label and the second it was said.
What that costs, measured rather than promised
The 12B model is 7 GB of weights, and it behaves like it. Measured on the real analysis path: memory peaks between 7.0 and 8.4 GB while a meeting is being analysed, then drops to roughly 180 MB about three minutes after the model goes idle. The download lands inside your own recordings folder, so you can see it and delete it.
Two consequences you will notice. On-device analysis pauses while a recording is running, because the microphone matters more than the summary. And the local model reads less of a long transcript per pass than the cloud route does — Narada’s own settings copy calls the result “lighter quality than the cloud”.
What “offline” does not cover
Worth being precise, because the word gets stretched.
- Calendar sync needs the network. Google and Microsoft calendars are OAuth APIs; there is no local copy to read on a plane. macOS Calendar via EventKit reads locally, but whatever fills it does not. You can also skip calendars entirely — a recording starts from the menu bar in one click, with or without an event behind it.
- The models need one download each. Roughly 600 MB for speech, plus 3 GB or 7 GB for the language model, from Hugging Face. After that, nothing.
- Apple Silicon only, for the language model. Intel Macs record and transcribe on-device perfectly well; the local summary does not exist there. On an Intel Mac, offline means Recorder-only.
- macOS 15.6 or newer, and Mac only. No Windows build, no iPhone app.
- Offline is a mode, not the whole app. Pro AI and bring-your-own-key both send transcript text to OpenAI after you explicitly consent, and Pro AI also sends a per-install device id to Narada’s proxy. Audio never leaves in any mode. If you want nothing to leave, the two modes to pick are Recorder-only and On-device AI.
Two things it will not do for you, on a plane or anywhere
No live transcription. Transcription runs after the recording stops, not during it. If you want words appearing on screen while someone talks, this is the wrong app.
No speaker names, and no speaker separation at all for in-person recordings. The call’s audio is split into Speaker 1, Speaker 2 and so on, and the app refuses to guess who they are. Your own lines are attributed to you only because your microphone is a separate file — so the meeting you record around a table on the plane comes back as one voice. Transcription covers 25 European languages, and not equally: the engine’s published benchmark runs from about 4% word error rate on Italian to about 37% on Greek.
One more thing that has nothing to do with the network. macOS audio can fail mid-recording — a device reshaped under a live capture, an overloaded audio daemon, a microphone that quietly stops delivering. Narada compares each finished recording against itself and marks it Incomplete with the reason. It cannot recover audio that was never written, and no local model helps with that.
And nothing about running offline makes a recording lawful. Rules differ by country and by state, and no app tells the room on your behalf — Narada shows a reminder on your own screen and a menu-bar indicator, and that is the extent of it. Say out loud at the start that you are recording.
Sources
All checked 4 August 2026.
- Granola: granola.ai/security (cloud processing, and why), docs.granola.ai — transcription (no audio saved), docs.granola.ai — network troubleshooting (Deepgram and AssemblyAI WebSocket endpoints).
- Otter: help.otter.ai — Recording Best Practices: Pause, Resume and Upload (offline capture, under a mobile-only heading), help.otter.ai — Best Practices for in-person recordings (stable connection required, 512 kbps uplink), help.otter.ai — Transcription processing time FAQ (offline recording processed after reconnecting).
- MacWhisper: docs.macwhisper.com — keeping transcriptions private (local by default; Ollama and LM Studio as the offline AI exception), docs.macwhisper.com — record meetings (meeting detection notifies, you click Record).
- Superwhisper: superwhisper.com/docs/models/language — defines local vs cloud, then lists language models from Superwhisper, Anthropic, OpenAI and Groq only, every one marked Pro; superwhisper.com/docs/get-started/sw-pro (local voice models are a Pro feature).
- Meetily: meetily.ai (100% on-device claim, Whisper.cpp, local Ollama summaries, minimum requirements).
- Narada: model tiers, RAM gates, memory measurements and the macOS 15.6 minimum from its own build and App Store listing. Word error rates per language from the FLEURS benchmark table published with FluidAudio 0.13.4, the transcription package Narada ships. https://raw.githubusercontent.com/FluidInference/FluidAudio/main/Documentation/Benchmarks.md
Questions people ask about this
Does Otter.ai work offline?
Not on a Mac. Otter’s help centre says “Otter requires a stable internet connection to provide real-time transcription”, and the only offline capture it documents is on the phone — the mobile app “can capture audio locally on your device when you’re offline”, then uploads when you reconnect. No equivalent path is documented for the Mac desktop app.
Can I get a meeting summary on a Mac with no internet at all?
Yes, if the app runs a language model locally. Narada ships one in its On-device AI mode on Apple Silicon. MacWhisper and Meetily can do it if you install and run your own local model server such as Ollama or LM Studio.
Which Macs can run Narada’s on-device summaries?
Apple Silicon with at least 8 GB of RAM, which gets the smaller 4B model; 16 GB or more gets the 12B one. Intel Macs run the recorder and the on-device transcription fine, but the local language model is not available on them at all.
Are on-device summaries as good as the cloud ones?
No, and Narada’s own settings screen says so — “lighter quality than the cloud”. The local model also sees less of a long transcript per pass than the cloud route does. For a summary you will read once, it is usually fine; for a two-hour multi-thread meeting the cloud route is sharper.
Does Narada need an internet connection for anything?
For calendar sync, for the initial model downloads, and for the cloud AI modes if you choose one. Recording, transcription, on-device summaries, and search and chat over your archive all work with the network off — and you can record with no calendar connected at all.
Try it on your next call
Narada records your Mac's audio and your microphone, transcribes on-device, and hands you the decisions and follow-ups. Free to use, no account, no bot in the meeting.