Quills

At a glance

Quills is a macOS menu-bar meeting recorder that saves your microphone and the Mac's system audio as two separate tracks, then transcribes both on-device with Parakeet and merges them into a transcript tagged me versus them. It is a downstream fork of Andrew Jones's Quill that adds multilingual models, a pause control and a local read-only MCP server so coding agents can search your transcripts.

Good fit

Mac users who already work in a terminal and want plain files from every call: CAF audio, a JSON transcript and a Markdown copy in a dated folder. It suits people who want Claude Code, Codex or OpenCode to look up past meetings through MCP, or who want an on_stop shell hook to start their own summarizing or filing step.

Consider something else when

Skip it if you want an installable app with summaries, notes or speaker names. Quills ships no app bundle or release, so you build and code-sign it yourself, and it only separates you from everyone else on the call. Skip it too if you need Whisper models, Windows or Linux, or a macOS version older than 15.

Start here

Clone the repository, run swift build -c release, and sign the binary with an Apple Development or self-signed code-signing identity before installing it to /usr/local/bin. Run quills models --download, then quills doctor. Click the feather icon, choose Start recording, grant microphone and System Audio Recording access, and stop to get a transcript in ~/Recordings.

Cost and requirements

The code is MIT-licensed and the README lists no paid service. It needs macOS 15 or later, the Swift toolchain to build, and a code-signing identity; Apple Silicon is recommended for transcription speed. The default Parakeet v3 model is about a 600 MB first download, after which transcription runs offline. Any MCP client you connect may send transcripts to its own hosted model.

Limits to know

The system-audio tap records everything the Mac plays, including notification sounds and music. Diarization is only mic versus system, and the echo filter is described as conservative, so speaker playback can still appear twice; mic_voice_processing helps when using speakers. An unsigned build asks for permissions on every recording. Sessions under 10 seconds are discarded. Parakeet v3 covers 25 European languages plus Japanese; the other two models are English only.

Related work

• Daisy, a Mac recorder that transcribes on-device with Whisper and exposes notes over a local MCP server (alternative)
• Meetily, a local meeting assistant using Whisper or Parakeet with Ollama summaries (alternative)
• Pluto, an open-source on-device meeting assistant for Apple Silicon Macs (alternative)

Still unclear

• How accurate the merged transcripts are on real multi-person calls; the README calls its model labels selection guidance rather than a measured ranking.