No. Narada uses macOS's own audio-tap API to capture what your Mac is playing, plus your microphone, as two separate tracks. Nothing joins the participant list, nothing is admitted to the call, and there's no BlackHole or Loopback-style virtual device to set up. It works with Zoom, Google Meet, Teams, Webex, Slack huddles — and with calls that aren't in a meeting app at all.

The audio file stays in a folder you pick on your Mac. The app does not upload raw audio. If you use Pro AI, transcripts and summaries are processed via our cloud — but you can switch to your own OpenAI key or a fully local LLM in Settings.

You keep every recording, transcript, and summary. They're plain files. The app keeps recording on the Free tier — you just stop getting AI-generated summaries, follow-ups, and chat.

Yes — we strongly recommend it, and in many jurisdictions it's required by law. Recording rules vary widely, and it's your responsibility to comply with the laws that apply to you and your participants. Narada keeps a reminder in the recording window every time you record, so it doesn't slip your mind.

Transcription runs on your Mac with Parakeet TDT v3, an on-device speech model, and quality depends heavily on the language. On the FLEURS benchmark its published word error rate ranges from about 4% for Italian and 5% for English to the mid-20s for Slovenian, Latvian, Lithuanian and Maltese, and 37% for Greek. English, German, French, Spanish, Italian, Dutch, Polish, Russian and Ukrainian are all in single digits or close to it. Anyone quoting you one accuracy number for every language is guessing.

Practical notes: your own voice is captured on a separate track from the call's audio, so cross-talk hurts less than it would in a single mixed recording, and there is no network round-trip to lose.

Yes — click the menu-bar icon and hit Record. Narada captures the microphone; there is no system audio to capture in a room. You get the same transcript, summary, decisions and follow-ups afterwards.

Two honest caveats. Speaker separation works on the call audio, not on a room recorded through one microphone, so an in-person transcript is not split by person. And because there is no call to listen to, the automatic stop can't tell that the conversation is still going — a room recording gets a fixed grace period past the calendar end rather than waiting for silence.

Yes, in two of the four modes. Recording and transcription are on-device, and On-device AI keeps the summaries there too — after the one-time model download, that combination is fully offline. Pro AI and a bring-your-own OpenAI key both need the network, because the model is somewhere else. Calendar sync obviously needs the network too, but you can record without a calendar at all.

No. Narada captures the other side of the call through a CoreAudio process tap — an audio API — not through screen capture. It asks for Microphone and System Audio Recording, and nothing else is required. Calendar access is optional; skip it and you still get manual recording.

Partly, and it's worth being precise. The call's audio is split into distinct voices on-device, so you get Speaker 1, Speaker 2 and so on with their lines separated. Narada does not guess which voice is which person — it has no voiceprints and doesn't build any. Your own lines are named, for a structural reason: your microphone is recorded as its own track.

You get told. macOS audio can fail mid-recording — a device changes underneath you, a driver stalls — and the worst version of that is a summary of half a meeting that reads like a summary of a short meeting. So Narada checks every finished recording against itself, comparing the audio it submitted with what actually landed in the file, and marks the meeting Incomplete in the list and on the detail screen when they disagree, saying which stream stopped and roughly how much is missing. It cannot recreate lost audio. It refuses to pretend.

The app is a universal binary, so recording, on-device transcription, Pro AI and your own API key all work on Intel. On-device AI is the exception — it runs on Apple Silicon's unified memory through Metal, and is not offered on Intel hardware.

All of them, because it doesn't integrate with any of them. Narada records what your Mac plays — Zoom, Google Meet, Teams, Webex, a Slack huddle, a Discord call, a phone call on speaker, a webinar you're only listening to. Nothing needs to expose an API, grant a bot access, or even know Narada exists.