quill is a fully local macOS meeting recorder and transcriber that went from nothing to 3,795 stars and 252 forks since the repository opened on 24 July 2026 — one of the sharpest three-week climbs on GitHub Trending this month. It is a single Swift binary with a menu-bar tray and no app bundle, sibling to the same author's parrot, and the entire pitch is in one line of the README: nothing ever leaves the machine.

The recording model is the interesting design decision. One click captures your microphone and all system audio as two separate tracks — mic.caf and system.caf. This is done for two concrete reasons rather than tidiness: speech models perform better on clean single-source audio, and splitting mic from system output gives free two-party diarization, me versus them, with no speaker-identification model in the loop at all. The CAF container is likewise chosen on purpose — unlike m4a it needs no finalization pass, so if the process dies mid-meeting everything already written to disk remains readable.

Transcription is built in, on-device, and automatic. The default engine is Parakeet TDT 0.6B v2 (English) through FluidAudio's Core ML port, which the author clocks at roughly 20 seconds per hour of audio on Apple Silicon. The ~600 MB of models download once on first transcription, and quill doctor reports whether they are already cached — so you are never waiting on a download right after an important meeting. Each track is transcribed separately, shifted by its recorded start offset so both share a single clock, then merged by timestamp. Jobs run in a serial queue, meaning a new recording can start while the previous one is still transcribing, and the filesystem is the queue: a session directory holding meta.json but no transcript.json is pending, so unfinished jobs simply resume on next launch. Failures append to that session's transcribe.log and never block later jobs.

Every session lands in ~/Recordings/<yyyy.MM.dd-HHmm>/ with both audio tracks, meta.json for timestamps and per-track offsets, a canonical transcript.json carrying engine provenance and timed speaker-tagged segments, a readable transcript.md, and the log. An optional on_stop hook in ~/.config/quill/config.json spawns a shell command with the session directory as its argument once the transcript is written — the intended seam for summarization, filing, or indexing. System audio is captured through a Core Audio process tap via a private aggregate device, so there is no virtual audio device and no kernel extension to install.

The constraints are narrow and worth stating. It is macOS 15 or newer only — the process-tap API requires it — with Apple Silicon recommended for transcription speed, so there is no Windows or Linux story whatsoever. The default engine is English-only; a WhisperKit large-v3-turbo fallback for re-transcription is described as planned, not shipped. A global tap records everything the Mac plays, notification dings and music included, which the README flags as a genuine gotcha. And with the newest commit dated 30 July 2026 against 23 open issues, the repository has been quiet for about two weeks while attention has been at its peak.