Your meetings, transcribed entirely on your Mac.

Why it exists

Cloud meeting AI means uploading the raw audio of every conversation to someone else's servers. For anything confidential, that's a non-starter.

So I built the opposite: transcription that runs entirely on-device, on open models, with nothing phoning home. And the transcripts aren't just notes — they're private context my own AI agents can draw on for total recall, without renting my memory to anyone. parley covers calls and meetings; mailrag covers email. Independent tools; my agents know about both and use what fits.

Why it's built this way

Meetings are sensitive. Most of what's said in them shouldn't be uploaded to anyone's servers, and for confidential work that rules out every cloud notetaker.

• On-device and airgap-capable. The one that matters most. Nothing leaves your Mac, ever.
• Open-source and free. Read the code, trust the code. Donations welcome. :)
• Crash-resilient. It shouldn't fall over in the meeting you most needed it for.
• Tamper-proof. Signed transcripts and summaries you can actually trust (on the roadmap).
• Agent-queryable. Your own AI agents can search the record over MCP, on your machine (also coming).

What it does

How it works — the hard parts

• Dual-stream capture (the core constraint). macOS exposes no API for a pre-mixed mic + system stream — verified through the macOS 26 SDK. So the app runs two independent capture streams — your mic and the system output — and treats "local" and "remote" as first-class. That's what makes reliable speaker separation possible.
• Capturing the calls a screen recorder can't see. ScreenCaptureKit is the default for system audio, but it misses Continuity (iPhone) calls and some VoIP audio. An optional Core Audio process tap grabs the system output directly, so phone and app calls land in the transcript too — switchable in Settings, off by default.
• Echo / mic-bleed removal. A triple-confirmed gate removes the far-end voice bleeding into your mic: >50% temporal overlap and >70% word overlap and >0.8 speaker-embedding cosine. Across 7 real recordings it caught 22% more far-end bleed than the heuristic it replaced (158 vs 129 segments), with zero false positives.
• Cross-chunk speaker reconciliation. Audio is chunked and transcribed in parallel; per-chunk speaker IDs are merged into one global identity via greedy cosine matching on embeddings.
• Crash-safe by design. Sentinel file + LaunchAgent restart + multi-segment stitching mean a crash mid-meeting costs ~300–800 ms, not your recording.
• Chunked over streaming — a deliberate accuracy call. Streaming ASR and diarization trade accuracy for low latency. Parley processes complete chunks in the background instead, so transcript and speaker labels come out as good as a full offline run, and only the final chunk waits for you to stop.
• On-device ML across the Neural Engine and GPU. Parakeet ASR and Silero VAD run on the Neural Engine; pyannote/WeSpeaker/VBx diarization runs through CoreML across the Neural Engine and GPU.

Speed

Measured on an M5 Pro (release build) with the bundled harness (tools/engine-benchmark), on a 4-minute AMI clip.

Chunks process in the background during the meeting, so the number you actually feel is the last row: at ~62× end-to-end, the final chunk (≤30 min by default, configurable) finishes in well under a minute after you stop, whatever the meeting's length. Apple SpeechAnalyzer isn't benchmarked here — it needs a per-language model and errored on this clip.

Open models, on-device

Get started: git clone https://github.com/fmasi/parley.git && cd parley && bash package_app.sh --install