Never forget to hit record.
Murmurly starts recording when your call starts and stops when it ends. Any app, any call. Speech is recognized on your Mac — nothing is uploaded.
No account. No subscription. No analytics.
Calls
It records on its own, so you don't have to remember.
When any app takes the microphone — Zoom, Google Meet in a browser, Telegram, FaceTime — recording starts. It stops 30 seconds after the call ends.
A call is detected by the microphone, not by a list of “supported services”, so it works with all of them. Don't want an app recorded? Exclude it with one command from the menu.
Privacy
Your voice stays on your Mac.
Speech recognition runs on-device. Murmurly never sends your recordings or transcripts anywhere. It goes online for two things only: downloading a model when you ask, and checking for updates.
~/Library/Application Support/Murmurly/Dictation
Talk instead of type.
Press ⌥ Space, speak, press it again — the text appears where your cursor is, in any app. Or double-tap ⌘ or ⌥. The shortcut is yours to change.
Nothing gets lost: every dictation is in a searchable history. If pasting fails, the text is there. If recognition fails, the audio is kept.
Transcripts
You and them, on separate tracks.
Your microphone and the call audio are recorded apart, on one clock. The transcript marks each line “Me” or “Others”, merges both sides by time, and removes the echo when your mic hears the speakers.
Right after the call you get transcript.md with timestamps, plus two lossless tracks (ALAC, 24-bit).
Models & languages
Two engines. Switch on the fly.
Both run offline. Models download from Hugging Face only when you ask, and switching needs no restart.
Fast. 25 European languages, including English and Russian.
About 100 languages. For everything Parakeet doesn't cover.
Interface in 12 languages. Transcribe existing files too: mp3, m4a, wav, mp4, mov.
Roadmap
Where this is going.
The promise stays the same at every step: your audio and text stay on your Mac unless you explicitly choose otherwise.
Alpha and first testers
This site, a handful of testers, and polishing what exists until it matches the promises.
Smart text processing
Modes for dictated and transcribed text: note, email, list, speech clean-up. Fully local, bring-your-own-key to Claude or another service, or a paid tier where we cover tokens. Cloud only by explicit choice, clearly marked.
Calendar integration
Connect your calendar so recordings know which meeting they belong to.
New engines and languages
Engines for languages Parakeet and Whisper miss (e.g. Indian languages). Streaming models that produce text while you speak — live captions.
Tools for AI agents
An MCP server and a CLI: agents like Claude can search history and transcripts, start and stop recording, and pull a meeting’s text straight into their work.
iPhone and Android
Mobile versions on the same principle: recognition on the device.
Try the alpha.
Mac with Apple silicon, macOS 15 or later. Free, no account.
The alpha is not notarized by Apple yet, so macOS asks once before the first launch. How to open it