pi-meeting-copilot

Live meeting transcription copilot for pi. Tracks a transcript file in real time, giving pi meeting context for questions, note-taking, and summarization.

Install

pi install git:github.com/samfoy/pi-meeting-copilot

This gives you:

• meeting_copilot tool — start/stop tracking, read incremental updates, read full transcript
• /meeting command — quick shortcuts for start/stop/status
• Skill file so the LLM knows when and how to use it

Usage

> I'm joining a standup, start the meeting copilot
# pi calls meeting_copilot(action: "start")

> what's been discussed so far?
# pi calls meeting_copilot(action: "read_new") and summarizes

> meeting's over, give me notes
# pi calls meeting_copilot(action: "read_all") and produces structured notes

Or use the command shortcut:

/meeting start
/meeting stop
/meeting status

Feeding It a Transcript

The extension reads a plain text file that grows over time — one line per utterance. It doesn't capture audio itself. Anything that appends text to a file works.

Default path: /tmp/live-transcript.txt

Override with MEETING_TRANSCRIPT_PATH env var or pass path to the tool.

Example Sources

whisper.cpp (local, Mac/Linux — best for real-time):

brew install whisper-cpp blackhole-2ch  # macOS
whisper-stream --model ~/.whisper-models/ggml-medium.en.bin \
  --capture 0 --file /tmp/live-transcript.txt \
  --step 5000 --length 10000 --vad-thold 0.8

Use BlackHole to route system audio (Zoom, Meet, etc.) to whisper as input.

Otter.ai / Zoom / Teams (cloud transcription): Export or copy the transcript to a file. For live use, some tools let you sync captions to a local file.

Any speech-to-text CLI that writes to stdout:

my-stt-tool --input mic >> /tmp/live-transcript.txt

Remote machine? If your transcription runs on a laptop but pi runs on a remote dev box, sync the file:

# From laptop, sync every 3 seconds
watch -n3 scp ~/transcript.txt devbox:/tmp/live-transcript.txt

# Or stream it
tail -f ~/transcript.txt | ssh devbox "cat >> /tmp/live-transcript.txt"

Format

No special format required. The copilot reads plain text, one line at a time. Timestamps are nice but optional:

[09:01:15] Let's start with the status update
[09:01:22] The deploy went out yesterday, no issues
we can also handle lines without timestamps

Configuration

| Env Var | Default | Description | |---------|---------|-------------| | MEETING_TRANSCRIPT_PATH | /tmp/live-transcript.txt | Where pi reads the transcript |

How It Works

1. A transcription source appends lines to a text file
2. meeting_copilot tracks your read position — read_new returns only lines since your last check
3. State persists across pi turns via session entries, so position survives context compaction
4. At meeting end, read_all gets the full transcript for summarization

Files

pi-meeting-copilot/
├── package.json              # Pi package manifest
├── src/
│   └── index.ts              # Extension — meeting_copilot tool + /meeting command
└── skills/
    └── meeting-copilot/
        └── SKILL.md          # Skill file for LLM context

License

MIT