---

Every call recorder wants to join your meeting as a bot, or ship your audio to someone's cloud. nastro does neither: it taps macOS system audio directly (CoreAudio process tap), writes a local file, and transcribes it on your machine — only when you ask.

nastro                      # TUI: browse, record, transcribe
nastro record               # headless recording, Ctrl+C = save
nastro records --plain      # scriptable list (tab-separated)
nastro transcribe last      # on-device transcript (diarization opt-in)

Why nastro

• No bot in the call. Capture happens at the OS level. Nothing joins your meeting, nothing is visible to other participants.
• No cloud. Audio and transcripts never leave your machine. Transcription runs locally (whisper.cpp or WhisperX).
• Who said what, one flag away. Set transcriber = "whisperx" for diarized output: [SPEAKER_00], [SPEAKER_01] — grouped by speaking turn, with SRT timestamps.
• Never lose a call. Audio is written incrementally: the file is valid even after a crash or kill -9. Transcripts are written atomically: an interrupted run never destroys an existing one. Deletes go to the Trash.
• Terminal-native. A fast single-binary TUI that inherits your terminal theme, with a headless CLI twin for scripting. Paths are OSC 8 hyperlinks — cmd+click opens Finder.

Quickstart

brew install scaccogatto/tap/nastro
nastro

Or from source:

git clone https://github.com/scaccogatto/nastro
cd nastro && make build
./bin/nastro

Press r, name the recording (or just hit enter), and talk. s stops and saves. t transcribes.

On first recording, macOS will ask for Screen & System Audio Recording permission — the grant goes to your terminal app.

Transcription

nastro is plug-and-play about its dependencies: it finds whisper-cli/whisperx even outside your PATH, and offers to install what's missing (via brew/uv, always with explicit confirmation).

The default backend is whisper.cpp, running on Metal: fast, no accounts. nastro installs it all itself, with confirmation — brew install whisper-cpp, then the ggml model, downloaded with progress on first use.

Want who-said-what? Set transcriber = "whisperx" for diarized, speaker-labeled output ([SPEAKER_00], [SPEAKER_01], grouped by speaking turn). It needs one manual step nastro can't do for you: a HuggingFace token with the pyannote model terms accepted — it's an access grant, not a download. Put it in your config as hf_token (or $HF_TOKEN). If anything is missing, the TUI walks you through it.

Keys

Configuration

~/.config/nastro/config.toml — everything has a sane default:

output_dir = "~/Recordings/nastro"
transcriber = "whisper-cli"     # or "whisperx" for speaker labels
whisper_model = "large-v3-turbo"
lang = "it"                     # transcription language
hf_token = ""                   # HuggingFace token for diarization

How it works

Two binaries, one job each:

• nastro (Go, Bubble Tea) — the TUI/CLI, orchestration, transcription pipeline
• nastro-tap (Swift) — audio capture via CoreAudio process tap + mic, incremental AAC writing, a red ● in the menu bar while recording

The filesystem is the source of truth: one directory per recording (audio.m4a, metadata.json, transcript.txt/.srt), no database. Everything the TUI does, the CLI does headless.

Requirements

• macOS 14.4+ (CoreAudio process tap API)
• Go + Swift toolchain to build from source

Roadmap

Separate mic/system tracks, optional video capture, markdown export. Design docs live in docs/.

License

MIT