Scribe

Your meetings, transcribed on your Mac.

Scribe sits in the menu bar, notices when a call starts, and records it if you say so. The transcript, with speakers and timestamps, is on your Mac a few minutes after you hang up. Nothing is uploaded.

Download for macOS

Version 0.2609270609.0, released 27 September 2026. 28.7 MB zip, notarized by Apple. SHA-256

macOS 15 or later on Apple silicon.

Product sync 47 min Zoom Maya Tom Priya
  1. 14:02
    Maya

    Let's lock the launch date. If the notarized build clears tonight we can ship on Thursday.

  2. 14:11
    Tom

    Support wants the release notes a day earlier. I'll draft them from this recording.

  3. 14:24
    Priya

    Can the pricing change wait until after launch? Two things changing at once is hard to explain.

  4. 14:31
    Maya

    Agreed, pricing waits. Tom, send the notes round by Wednesday noon.

  5. 14:38
    Tom

    Will do. I'll pull the action items out of this transcript and post them after the call.

It asks before it records

When Zoom, Teams, Meet, Slack, FaceTime or another call app opens the microphone, a small chip appears under the menu bar. Press Record and it becomes that recording's transport. Dismiss it and Scribe stays out of the way until the next call.

Transcription runs on your Mac with a local speech model, and the finished transcript opens with speakers told apart, timestamps and search.

Dictate into any app

Hold the right Command key, speak, and let go. The words land in whichever field has focus, in Mail, Slack, a browser or a terminal. Double-tap to keep listening hands-free.

Dictation is offline and keeps no history. Audio stays in memory and is gone once the text is inserted.

Ask your assistant about a meeting

Connect ChatGPT, Claude or Claude Code once, and they can find recent meetings, search your transcripts and quote timestamped text for summaries and action items.

Transcripts stay on your Mac. Assistants read them only through a connection you approved, and you can disconnect any of them from Scribe at any time.

Built on open models

Scribe does not train or host a model of its own. It runs published, open speech models on the Neural Engine, converted to Core ML and pinned by checksum, so what transcribed your meeting is something you can read about, inspect and run yourself.

Together they are a one-time download of about 505 MB from Hugging Face. After that, Scribe works with the network off.

  1. Parakeet TDT 0.6B v3 CC BY 4.0 NVIDIA's speech recognition model turns the recording into words with timestamps, in 25 European languages, at many times real time on Apple silicon.
  2. pyannote segmentation 3.0 MIT Finds where speech starts and stops, and where two people talk over each other, so overlapping turns can be marked rather than merged.
  3. WeSpeaker embeddings with VBx clustering Apache 2.0 Give each stretch of speech a voiceprint and group the voiceprints into speakers, which is how the transcript tells Maya from Tom without being told who was there.

Core ML conversions by FluidInference (CC BY 4.0), run through the open-source FluidAudio library (Apache 2.0). Exact revisions and checksums are in Scribe's model manifest.

Recording, transcription and dictation all happen on your Mac. There is no account to create and no cloud to trust.

Once the models are installed, Scribe works with the network off. The only thing that ever leaves your Mac is an answer to an assistant you connected yourself.