A new Mac app called Trace transcribes meetings without sending audio to a server or adding a bot to the call. The catch is a hard platform tax: M series Macs only, no team features, and accuracy that depends on the room.
The dominant design of meeting transcription has a quiet cost: a stranger joins the call as a bot, the audio leaves the room, and the recording comes with a terms-of-service grant. A new Mac application called Trace is making a deliberate case that this is not the only way to do the job. The premise is that local software can now transcribe a meeting on the same laptop that hosts it, and that the tradeoffs of that choice are concentrated in exactly the people the product is not built for.
The architecture is specific enough to reason about. Trace records microphone and system audio on the Mac, transcribes with a local speech model, and lets the user flag key moments inline. The product page states that no audio, transcript, or metadata leaves the device. Speaker diarization runs locally using two open-source models that ship with the app: a Pyannote-based segmentation model and a WeSpeaker embedding model. The first launch pulls roughly 500 MB of those models from Hugging Face, and the vendor says that is the only required network request. From there the app runs sandboxed, with a single declared network entitlement for that one-time model fetch.
The cost of that design shows up in three places. The first is platform. Trace requires macOS 14 or later and Apple silicon, and is distributed through the Mac App Store. That excludes every Intel Mac, every Windows and Linux user, and most enterprise defaults. For a category that has spent two years chasing cross-platform reach, the choice is a deliberate narrowing: a tool for the audience that already has the hardware to run it, not a tool for everyone.
The second is collaboration, or its absence. There are no team accounts, no shared notebooks, no cross-device sync. Output is per-session folders containing mic.wav, system.wav, transcript.json, and transcript.md, browsable from Finder, with key moments flagged at their original timestamps. The product positions this as a virtue: no accounts, no servers, no API keys. It is a virtue. It is also a gap for the way most companies actually use meeting tools, where a shared record of a customer call or a manager's later review of what was said is the whole point.
The third is accuracy. On-device speech recognition has improved, and shipping open-source speaker models is a defensible choice, but the vendor's privacy policy does not promise uniform performance across accents, languages, or noisy rooms. Independent benchmarks are not in hand. For native English speakers in quiet rooms, the design is likely fine. For the long tail of accents and acoustics that real meetings produce, accuracy is a question the user has to answer in their own environment. Privacy and correctness are not the same property, and the product's strongest claim covers only the first.
There is a smaller tension inside the speaker-labeling story. The product page shows named speakers, including a placeholder for someone called Alex. The privacy policy describes labels as generated output, typically "Speaker 1, Speaker 2," without claiming that the app identifies real people. The difference matters: diarization groups who spoke when, while named labels would require either a voice profile or external information. The honest read is that Trace groups speakers and lets the user rename them, which is the conservative and correct interpretation.
Read as a single product, Trace is a small app with a sharp opinion. Read as a category signal, it is more pointed. The combination of on-device speech models, sandboxed entitlements, and Apple silicon throughput is what makes a no-cloud meeting AI technically possible in 2026. The fact that a small team is willing to ship the product without live collaboration, cross-platform support, or a server backend is a bet that this slice of the market prefers a private, single-user tool to a shared one. The platform tax, the collaboration tax, and the accuracy variance are the price of that bet, and they are the parts a reader should weigh before installing.