What is Plaudy?
Plaudy is an open-source macOS meeting recorder that runs 100% on-device.
One click, or fully hands-free, captures both sides of a call: your microphone and the Mac's system audio. It lands as a single, speaker-attributed transcript you can play back, search, and summarize with the AI you already pay for. Cloud recorders ask you to trade privacy for convenience; Plaudy was built so you don't have to choose.
Who said what, not a wall of text
On-device speaker diarization builds a per-speaker timeline: Me at 0:04, Speaker 1 at 0:09, each with its own words. It works offline, from the first recording, because the models ship inside the app.
Both sides of the call, natively
System audio comes in through the CoreAudio Process Tap built into macOS 14.4 and later. Nothing to install: no BlackHole, no virtual devices, no kernel extensions.
Your AI, not theirs
A local, read-only MCP bridge hands transcripts to your Claude, so summaries run on the account you already have. Prefer zero network? Point the same bridge at a local model.
Honest by design
The menu-bar icon changes shape into an ear whenever Plaudy listens, so recording is never invisible. Auto-capture is opt-in and never records your bare microphone on its own.
How does it work?
Every step below happens on your Mac; the network is never part of the pipeline.
Mic and system audio feed the same faithful capture and get mixed into one playable recording. Two on-device models then run side by side: diarization (sherpa-onnx) figures out who spoke, and ASR (Parakeet or Whisper, Metal-accelerated) figures out what they said. A merge step aligns the two into a speaker timeline. The last arrow, the AI summary, is the only one you opt into, and it goes to your AI.
How is Plaudy different from Plaud, Otter & Fireflies?
They upload your audio to their servers and process it there. Plaudy does the same work with models that run on your Mac, so the recording never leaves the machine.
| Question to ask | Plaudy | Plaud · Otter · Fireflies |
|---|---|---|
| Where does my audio go? | Stays on your Mac, always | Uploaded to their cloud |
| Do I need an account? | No account, no sign-in | Yes |
| Who said what? | On-device diarization, per-segment timeline | Yes, computed in their cloud |
| System-audio capture? | Native CoreAudio tap, no virtual driver | Hardware gadget, or a bot joins your call |
| Whose AI summarizes? | Yours: your Claude, your account, or a local model | Their LLM, on their servers |
| What does it cost? | Free, open source, MIT | Subscription |
| Can I audit the code? | Every line, on GitHub | Closed |
Honesty over flattery: if you need live transcription on Windows or Linux today, the open-source peer Meetily is excellent. Different bets, mutual respect. Plaudy bets on open on-device diarization, an agent-ready MCP bridge, and a Mac-native feel. Found an error in this table? Open a PR; we would rather be accurate than flattering.
Is it real, or another demo?
The core is built and was validated live, with real speech and real meetings. Here is exactly how.
- Dual-stream capture, validated with live speech. A real session with talk on both streams: the microphone labeled "Me", the system audio diarized into speakers, both merged into one timeline. When the mic re-hears the speakers, an echo de-dup step keeps one person from splitting into two.
- Diarization, validated on 1, 2 and 3-speaker recordings. Each speaker count separated correctly, silence handled, and the two ONNX runtimes (diarizer and ASR) running together without a crash.
- Auto-capture, validated end-to-end against external audio. With the app's own audio tap excluded from detection, an external sound started a session in about 1.4 seconds and the session finalized on its own when the audio stopped.
- 106 automated tests green: 102 in Rust, 4 on the MCP bridge. Every claim above is documented in the repo's handoff notes with file-and-line citations, so you can check the receipts instead of trusting this page.
Questions people actually ask
Is Plaudy really private?
Yes: capture, transcription and diarization never touch the network.
Recordings and the SQLite database live in ~/Library/Application Support and stay there. No telemetry, no Plaudy account, no Plaudy server. The MCP bridge speaks stdio only (no listener, no socket) and opens the database read-only. The only network step is the one you choose: asking your own AI to summarize a transcript, under your own account. Prefer zero network? Point the bridge at a local model.
Does Plaudy know who said what?
Yes. On-device diarization builds a per-speaker timeline with timestamps.
pyannote-segmentation and TitaNet run locally via sherpa-onnx. Your mic is labeled "Me"; everyone in the system audio becomes Speaker 1, Speaker 2, and so on. When your speakers play the call out loud and the mic re-hears it, an echo de-dup step keeps one person from splitting into two.
Do I need a virtual audio driver like BlackHole?
No. System audio is captured natively via the CoreAudio Process Tap on macOS 14.4 and later.
No virtual devices and no kernel extensions. It uses the audio-capture permission, not screen recording, so there is no purple banner in your menu bar.
How do AI summaries work without a cloud?
Plaudy ships a local MCP server that your Claude connects to on your machine.
Instead of bundling a weak local model or renting you theirs, Plaudy exposes your library (list, get, search) to the AI you already use. The transcript text goes to your AI under your account; audio and database never move. The bridge is model-agnostic: point it at a local LLM for a fully offline loop.
How much does it cost?
Nothing. Plaudy is free, open-source, MIT-licensed.
Clone the repo and build it: the documented setup needs neither Homebrew nor full Xcode, just Command Line Tools on Apple Silicon. A notarized .dmg is on the roadmap.
What's the catch?
It's early, and this page says so.
Plaudy is macOS-only (14.4+), there is no live transcription yet, auto-capture is opt-in and experimental, and today you build from source. The core (dual-stream capture, diarization, the MCP bridge) is built and validated live. Roadmap and dev onboarding are in the repo.