Call Copilot

One thing. Done really well. Record meetings. Save transcripts. Don't break audio.

Call Copilot is a terminal command (rec) that records your meeting audio and, when you stop it, transcribes locally and saves a clean markdown transcript. Everything runs on your machine — no cloud, no API keys, no account. Audio and transcripts never leave the computer.

$ rec setup         # one-time: verify macOS + grant capture permission
$ rec start         # starts recording, you keep working
$ rec stop          # stops, transcribes, saves markdown
$ rec list          # shows past recordings

How it captures audio (no BlackHole, no Multi-Output Device)

By default, rec captures both the microphone (your voice) AND system audio (other participants / anything apps play) — so a real meeting where you speak gets recorded, not just the audio coming out of your speakers. Both sources are tapped directly via Apple's Core Audio taps API (macOS 14.2+) using the audiotap library. There is:

• no virtual audio driver to install (no BlackHole),
• no Multi-Output Device to create in Audio MIDI Setup,
• no system output device to switch and restore,
• and therefore no silent-recording failure that the old driver-based approaches produce when routing breaks.

The mic and system streams are recorded as two separate WAVs (each at its own true rate), transcribed separately, and merged into one transcript with [Mic] / [System] labels so you can tell who said what. Want just one source? rec start --mic-only or --system-only.

Grant capture permission (one-time, per terminal app)

macOS ties capture permission to the app you run rec from (Terminal, iTerm, Warp, VS Code, …). You need two grants:

1. Microphone (your voice): System Settings → Privacy & Security → Microphone → enable your terminal app.
2. System audio (grouped under Screen Recording on macOS 14.2+): System Settings → Privacy & Security → Screen Recording → enable your terminal app, then quit and reopen it (required for the change to take effect).

Until mic is granted, rec start records system audio only and skips the mic (it tells you so). rec setup checks the current mic-permission status.

After rec stop (or Ctrl+C from rec start), the recording(s) are transcribed locally with faster-whisper (CPU, int8 — no API, no data leaves your machine) and written to a markdown transcript.

recording.wav ──> faster-whisper (local, free) ──> transcript.md

Requirements

• macOS 14.2 or later (Sonoma — the Core Audio process taps API rec records through landed in 14.2). Apple Silicon recommended. Intel works.
• Python 3.11+ only needed for the pipx install; the Homebrew formula brings its own.

No virtual audio driver (no BlackHole), no Multi-Output Device, no manual Audio MIDI Setup. The tap reads system output directly. rec checks the macOS version itself and refuses to run below 14.2 with a one-line message — you don't have to guess.

Install

Homebrew (recommended — no Python to manage):

brew install AnisurRahmann/tap/call-copilot

pipx (if you prefer a plain Python install, no Homebrew):

pipx install call-copilot

Either way you get a rec command on your PATH. Then jump to One-time setup.

Using rec on a machine below macOS 14.2 prints Error: rec requires macOS 14.2 or later ... and exits — no traceback, no mystery. See docs/HOMEBREW.md if you're setting up the tap yourself.

Recording consent & privacy

Call Copilot records audio from your meetings, which may include other participants. Recording laws vary by jurisdiction and many require the consent of everyone being recorded (one-party vs. all-party consent). Obtaining that consent is your responsibility — check your local laws and your organization's policy before recording.

On the technical side, your data stays put:

• Transcription runs 100% locally with faster-whisper (CPU, int8) — no API, no cloud.
• Audio and transcripts never leave your machine. The only network access is the first run of a given Whisper model, which downloads its weights from Hugging Face; after that it's fully offline.
• Recordings live under ~/.local/share/rec/sessions/ (XDG data home), outside any repo.

One-time setup

rec setup

Verifies your macOS version, that the audiotap library + its bundled dylib load, saves your config, and tells you about the one-time capture permission prompt.

Recording a meeting

rec start                          # mic + system (default); Ctrl+C to stop & transcribe
rec start --system-only            # just what apps play (not your voice)
rec start --mic-only               # just your voice (not system audio)
rec list                           # browse past sessions
rec transcribe 2026-07-27_14-30-00 --model medium    # re-transcribe at higher quality
rec diagnose 2026-07-27_14-30-00                      # bundle debug info for an AI agent

By default rec start records both your microphone and system audio; the transcript labels each line [Mic] or [System]. Use --system-only / --mic-only to narrow.

rec start shows a live indicator while recording:

● REC  2026-07-28_14-30-00
elapsed 03:47   size 18.2 MB
press Ctrl+C to stop & transcribe

Press Ctrl+C when your meeting ends — it stops the recording and transcribes in the same command, then prints the transcript path. One command, start to finish.

Want the old background behavior instead? rec start --detach spawns the recorder and exits immediately; stop it later from another terminal with rec stop (or check progress with rec status).

Tip: if a transcript comes back empty, the recording was silent — nothing was playing, or the capture permission was revoked. rec start (and rec stop) warn you about this immediately, and rec diagnose <session> bundles the audio levels + logs.

Configuration

Stored at ~/.config/rec/config.json (XDG). Recordings live under ~/.local/share/rec/sessions/{id}/. Override the XDG roots with XDG_CONFIG_HOME / XDG_DATA_HOME if needed.

{
  "sample_rate": 16000,
  "channels": 1,
  "whisper_model": "base",
  "capture": "system",
  "sessions_dir": "~/.local/share/rec/sessions"
}

16 kHz mono float32 is Whisper's native input format — no resampling, smallest files (~7.5 MB/min), best transcription accuracy.

A note on sample rate (why your recordings play at the right speed)

audiotap's sample_rate parameter is not honored — the tap always delivers audio at your output device's native rate (typically 48 kHz), regardless of what we request. rec handles this transparently: the recorder measures the true capture rate when the tap starts and writes the WAV at that rate (so playback is the correct speed), and the transcriber resamples to 16 kHz before feeding Whisper (so transcription is accurate). You don't need to do anything; this section exists to explain the capture_sample_rate field in session.json (which may differ from the sample_rate in your config).

A note on VAD (voice activity detection)

Transcription runs without faster-whisper's Silero VAD pre-filter by default. The VAD is tuned for close-mic speech and aggressively rejects system-audio capture (speakers/headphones via a tap, which has a different character) — we've seen it discard 100% of a clearly-audible recording and return an empty transcript. Whisper's own no_speech_threshold handles silence adequately without that risk.

If you have clean close-mic input and want long silences skipped (faster, cleaner output), enable it per run:

rec stop --vad              # VAD on for this transcription
rec transcribe <id> --vad   # re-transcribe with VAD

Stack

Everything is free and open-source. Transcription runs 100% locally — no API cost, no data leaves the machine. (The first run of a given whisper model downloads its weights from Hugging Face; after that it's offline.)

Logging & debugging

Every activity — the audio tap lifecycle, chunk writes, transcription, formatting, each CLI decision — is logged. Logs flow to three destinations:

* WARNING by default: -v → INFO, -vv → DEBUG, --quiet → CRITICAL, REC_LOG_LEVEL=DEBUG.

Every command failure is logged at ERROR with the reason + exit code; unexpected crashes include the full traceback.

Hand a session to an AI agent to debug

rec list                                 # find the session id
rec diagnose 2026-07-27_14-30-00        # → writes sessions/<id>/diagnose.md
rec diagnose 2026-07-27_14-30-00 --stdout | <your-ai-tool>   # pipe straight to an agent

The bundle contains the session metadata, the daemon's recorder.log, the global-log lines tagged with that session, the transcript if it exists, and the config — followed by a debugging checklist. (rec diagnose accepts a unique prefix too, e.g. 2026-07-27.)

Note on the recording loop: the audio callback runs on Core Audio's real-time thread and deliberately logs nothing — it copies each chunk to a queue and a writer thread does the disk I/O. The daemon logs only at tap start/stop and on signals.

Development

You only need this if you're hacking on rec itself. End users should use the Install commands above.

git clone https://github.com/AnisurRahmann/call-copilot.git && cd call-copilot
make install          # creates .venv and installs `rec` in editable mode + dev deps
make test             # unit tests (offline — no audio device or model download needed)
make run              # rec start, using the dev venv's rec

make install requires Python 3.11+ on your PATH (it pins python3.11 for the venv). For everything else — releasing, the Homebrew tap, code style — see CONTRIBUTING.md and docs/HOMEBREW.md.

Contributing

Issues and pull requests are welcome at github.com/AnisurRahmann/call-copilot. Please read CONTRIBUTING.md before opening a PR, and report security issues privately per SECURITY.md.

What this is NOT

• Not a real-time transcription tool (transcription happens after you stop).
• Not a meeting summarizer (just the transcript).
• Not a Zoom/Meet plugin (it captures system audio generically).
• Not cross-platform (macOS 14.2+ only).

License

MIT — see LICENSE. © Anisur Rahman.