Your meetings stay on your Mac. Your own assistant can work with them.

Vedivox records your calls, transcribes them and works out who said what, all on your Mac. Claude Code or any other MCP client can then search them, follow a call while it is still going and write the summary. No account, no cloud, no bot in the call.

Get Vedivox
  • Free software, GPL-3.0-or-later
  • macOS 14.4+, Apple Silicon
  • 25 European languages
A meeting in Vedivox: the player with a waveform coloured by speaker, the transcript line by line with each speaker’s name, and a description on the right A meeting in Vedivox: the player with a waveform coloured by speaker, the transcript line by line with each speaker’s name, and a description on the right

Vedivox sends your meetings nowhere

Recording, transcription and speaker recognition run on the Mac’s Neural Engine. Audio, transcripts, voice profiles and the database stay in ~/Library/Application Support/Vedivox/.

There is no account, no telemetry, no analytics and no crash reporting. Vedivox calls no cloud or AI service, and nothing in it listens on the network. It records what your Mac plays and hears, so no bot joins the call.

The app is not sandboxed, so this is a property of the code, not something macOS enforces. The source is on GitHub.

In many places, recording a conversation needs the consent of everyone in it. Vedivox does not tell the other people that it is recording. Asking them is up to you.

What does touch the network

  • On first launch the speech models, about 1.2 GB, are downloaded once from Hugging Face. After that Vedivox works offline.
  • Once a day it asks GitHub whether a newer version is out. Nothing about you is sent, and you can turn it off in Settings.
  • Importing from a link runs yt-dlp, which you install yourself and which contacts the video site.
  • Whatever your assistant reads through vedivox goes to the model that assistant uses. That is your choice of assistant, not Vedivox.

Ask your assistant what was just said

While a meeting records, Vedivox keeps a live text of it. The vedivox command gives your assistant that text and every past meeting: it can answer a question mid-call, find what someone said last month, rename a speaker or write a meeting’s title and description.

Vedivox itself calls no language model, so without an assistant there are transcripts but no summaries. The release zip includes three Claude Code skills, /vedivox, /vedivox-live and /vedivox-enrich, and other MCP clients can use the vedivox-mcp server.

Live text of a recording in progress: lines from Clara Meyer, Speaker 2 and You Live text of a recording in progress: lines from Clara Meyer, Speaker 2 and You
vedivox live --last 5m
The last five minutes of the call in progress
vedivox search "beta list"
Every transcript line that mentions it
vedivox transcript 3f2a --json
A whole transcript with speakers and times, for a script
vedivox speaker rename …
Name a voice; later meetings use the name

Voices it has met get their names

Vedivox records two tracks at once: your microphone, and the system audio of the call. Zoom, Google Meet, Microsoft Teams, Yandex Telemost or anything else that plays sound.

A new voice shows up as Speaker N until you name it; after that Vedivox recognizes it in later meetings. Move a line to the right person or merge two speakers, and reprocessing keeps your fixes.

  • Your microphone Your side of the call, labelled as you
  • The call Split into voices, one speaker each
The Speakers screen: voices Vedivox has learned, with the meetings each person was in The Speakers screen: voices Vedivox has learned, with the meetings each person was in

What else it does

Dictation with a hotkey

Press ⌘⇧2 in any app and speak; the text is typed where your cursor is. Music pauses while you talk.

Notices calls

Zoom, Teams, Google Meet and Telemost calls are detected, and Vedivox reminds you to record. Starting and stopping on its own is a setting, off by default.

Keeps parts out

Press ⌃⇧⌘1 and what is said from then on stays out of the transcript until you press it again.

Imports recordings

Drag an audio file onto the window and it gets the same transcript and speakers as a live recording.

Trims and splits

Cut the chatter at the start of a call, or split one recording into two meetings.

Learns your words

A plain vocabulary file fixes the names and jargon speech recognition gets wrong, in transcripts and in dictation.

Get Vedivox

With Homebrew:

brew install --cask dimstunt/tap/vedivox

That installs the app, the vedivox command and the MCP server, and prints the two commands that install the Claude Code skills as a plugin.

Or download the latest release as a zip and run bash install.sh in the unpacked folder: it installs all of that, skills included. INSTALL.md walks through every step.

Vedivox is free software under the GNU GPL, version 3 or later, and its source is on GitHub. It is pre-1.0, built for one person’s daily use and shared as is; database migrations may still break compatibility between versions.

Requirements
macOS 14.4 or later on Apple Silicon. Internet on first launch for the model download, and about 3 GB of free disk for the models and their caches.
Permissions it asks for
  • Microphone, on the first recording.
  • Screen & System Audio Recording, to hear the other side of a call.
  • Optional: Accessibility for typing dictated text, Calendar to link meetings to events, Automation to notice a call in a browser tab.
Removing it
Your library is in ~/Library/Application Support/Vedivox/; the speech models are in ~/Library/Application Support/FluidAudio/Models. Delete both by hand.
Built on
FluidAudio, NVIDIA’s Parakeet speech model, and the pyannote and WeSpeaker speaker models.
Supported languages

Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian. Other languages are not supported.