pi-notetaker

Built for cmd0 — a desktop AI agent that lives in your system tray.

A pi extension for meeting notes. Record audio, transcribe it locally with Whisper, and generate structured summaries using the active pi LLM — all offline, all local.

Works with any pi agent (standalone pi, cmd0, or custom SDK apps).

Quick Install

# Install from npm
pi install npm:@dvrosalesm/pi-notetaker

# Then inside a pi session, run:
/meeting setup

Install from source

# Copy or symlink this directory into pi's extensions
cp -r pi-notetaker ~/.pi/agent/extensions/pi-notetaker

# Then inside a pi session, run:
/meeting setup

The /meeting setup command (or the meeting_setup tool) will:

• Install sox via Homebrew (for audio recording)
• Install whisper-cpp via Homebrew (for local transcription)
• Download the whisper base.en model (~142MB) to ~/.pi/models/

Manual Install (if Homebrew is unavailable)

# Recording — install one of:
brew install sox          # provides the `rec` command
brew install ffmpeg       # alternative recorder

# Transcription — install one of:
brew install whisper-cpp  # provides `whisper-cli`
# OR run a local OpenAI-compatible whisper API server

# Whisper model — download to ~/.pi/models/
mkdir -p ~/.pi/models
curl -L -o ~/.pi/models/ggml-base.en.bin \
  https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin

Prerequisites

| Dependency | Purpose | Install | |---|---|---| | sox or ffmpeg | Audio recording | brew install sox | | whisper-cpp | Local speech-to-text | brew install whisper-cpp | | Whisper model (.bin) | Whisper weights | Auto-downloaded by /meeting setup | | Homebrew (optional) | Auto-install above | brew.sh |

Platforms: macOS and Linux. Windows is not supported.

For system audio capture (not just mic), install BlackHole and configure a Multi-Output Device in Audio MIDI Setup, then set the device:

/meeting config device "BlackHole 2ch"

Usage

Commands

| Command | Description | |---|---| | /meeting start [name] | Start recording audio | | /meeting stop | Stop recording, transcribe & summarize | | /meeting list | List past meetings | | /meeting view <id> | View meeting summary | | /meeting transcript <id> | View raw transcript | | /meeting status | Check if currently recording | | /meeting setup | Check & install dependencies | | /meeting config [key val] | Show or set configuration |

LLM-Callable Tools

These tools are registered so the LLM can manage meetings conversationally:

| Tool | Description | |---|---| | meeting_setup | Check & auto-install all dependencies | | meeting_start | Start recording (accepts optional name) | | meeting_stop | Stop, transcribe, and summarize | | meeting_status | Check recording state | | meeting_list | List past meetings | | meeting_view | View a meeting's summary or transcript |

Agent Integration

To have an AI agent install this extension and set it up programmatically:

1. Copy this directory into ~/.pi/agent/extensions/pi-notetaker/
2. Call the meeting_setup tool with { "install": true } — it will install all system dependencies and download the whisper model
3. Use meeting_start / meeting_stop to record and process meetings

Configuration

Config is stored at ~/.pi/meetings-config.json (or ~/.cmd0/meetings-config.json if running inside cmd0). Set values with /meeting config <key> <value>.

| Key | Description | Default | |---|---|---| | device | Audio input device name | System default mic | | recordingBackend | "sox" or "ffmpeg" | Auto-detect | | whisperBackend | "cli" or "api" | Auto-detect | | whisperBin | Path to whisper binary | Auto-detect | | whisperModel | Path to .bin model file | Auto-detect in ~/.pi/models/ | | whisperApiUrl | Local whisper API URL | http://localhost:8080 | | language | Transcription language | "en" | | sampleRate | Recording sample rate (Hz) | 16000 |

Environment variables WHISPER_BIN, WHISPER_MODEL, and WHISPER_API_URL are also respected.

How It Works

/meeting start "Weekly Standup"
  |
  v
[sox/ffmpeg records audio -> ~/.pi/meetings/<id>/audio.wav]
  |
/meeting stop
  |
  v
[whisper-cpp transcribes audio -> transcript.txt]
  |
  v
[pi's active LLM summarizes transcript -> summary.md]
  |
  v
Summary injected into conversation context

1. Record — Spawns rec (sox) or ffmpeg to capture audio at 16kHz mono WAV
2. Transcribe — Runs whisper-cli locally against the audio file, or POSTs to a local OpenAI-compatible API
3. Summarize — Sends the transcript to pi's currently active LLM (same model you're chatting with) with a structured prompt
4. Store — Everything saved to ~/.pi/meetings/<timestamp>/ with metadata

Data Storage

The extension auto-detects the data root:

• cmd0 users → ~/.cmd0/ (if it exists)
• All other pi agents → ~/.pi/

~/.pi/                         # or ~/.cmd0/
  meetings/
    2026-03-30_14-30-00_a1b2c3/
      audio.wav                # Raw recording
      transcript.txt           # Whisper output
      summary.md               # LLM-generated summary
      metadata.json            # Name, timestamps, flags
  meetings-config.json         # Extension settings
  models/
    ggml-base.en.bin           # Whisper model weights

Extension Structure

pi-notetaker/
  package.json           # pi extension manifest
  README.md              # This file
  src/
    index.ts             # Extension entry — commands, tools, events
    recorder.ts          # Audio capture (sox / ffmpeg)
    transcriber.ts       # Local Whisper (CLI / API)
    summarizer.ts        # LLM summarization via pi's active model