echoloop

name	echoloop
description	Real-time AI meeting copilot. Dual audio capture (system + mic), live transcription (faster-whisper local or Deepgram cloud), LLM coaching loop (Claude/GPT every ~35s), always-on-top overlay, session logging with markdown export. Use when the user says "meeting recap", "meeting copilot", "live transcription", "meeting insights", "coaching during meeting", or "record this call".
version	1.0.0
author	Pete Woodbridge (EchoLoop)
tags	["meetings","transcription","audio","coaching","real-time","whisper","deepgram"]
env_vars	["ANTHROPIC_API_KEY","OPENAI_API_KEY","DEEPGRAM_API_KEY","ECHOLOOP_SYSTEM_DEVICE","ECHOLOOP_MIC_DEVICE"]

EchoLoop — Real-Time Meeting Copilot

Live meeting transcription + AI coaching. Captures dual audio (system speakers + microphone), transcribes in real-time, and feeds rolling transcript to Claude or GPT for tactical insights every ~35 seconds.

When to Use This Skill

• Live meeting coaching: Real-time tactical advice during calls
• Meeting transcription: Dual-stream (you + them) labelled transcript
• Meeting recap: Session logs with markdown export
• Negotiation support: Reads the room and suggests responses
• Multi-user audio: Works with our VNC desktop + virtual audio cables

When Not to Use

• For pre-recorded audio/video processing — use ffmpeg-processing
• For generating podcasts from text — use notebooklm (audio overviews)
• For video production — use open-montage
• For browser-based meeting automation — use playwright

Architecture

┌───────────────┐     ┌──────────────┐     ┌──────────────┐     ┌────────────┐
│ Audio Capture  │────>│  Transcriber │────>│   EchoLoop   │────>│     UI     │
│ (2 threads)    │     │ (thread pool)│     │   Engine     │     │  (Tkinter) │
│                │     │              │     │  (asyncio)   │     │            │
│ system ──┐     │     │faster-whisper│     │              │     │ Always-on- │
│ mic ─────┤     │     │  or Deepgram │     │ Claude / GPT │     │ top overlay│
│          v     │     │              │     │ every ~35s   │     │            │
│  chunk queue   │     │  seg queue   │     │ + on silence │     │ insight log│
└───────────────┘     └──────────────┘     └──────────────┘     └────────────┘

Quick Start

# Install dependencies
pip install -r ~/.claude/skills/echoloop/requirements.txt

# Set API key (uses existing container keys)
export ANTHROPIC_API_KEY="$ANTHROPIC_API_KEY"  # Already set for devuser

# Configure audio devices (container VNC)
export ECHOLOOP_SYSTEM_DEVICE="pulse"     # PulseAudio virtual cable
export ECHOLOOP_MIC_DEVICE=""             # Default mic

# Launch on VNC Display :1
DISPLAY=:1 python3 -m echoloop

Container Audio Routing

Our container has PulseAudio on VNC Display :1. For meeting capture:

1. Virtual audio cable: Route system audio through PulseAudio monitor
2. Set device: ECHOLOOP_SYSTEM_DEVICE=<pulse monitor name>
3. Launch: DISPLAY=:1 python3 ~/.claude/skills/echoloop/main.py

For multi-user scenarios, each user can run their own EchoLoop instance with different audio routing.

Environment Variables

Variable	Default	Description
ANTHROPIC_API_KEY	(required)	Claude API key (already set in container)
OPENAI_API_KEY		GPT API key (for OpenAI provider)
DEEPGRAM_API_KEY		Deepgram key (for cloud transcription)
ECHOLOOP_LLM_PROVIDER	anthropic	LLM provider: anthropic or openai
ECHOLOOP_TRANSCRIBER	local	Transcription: local (faster-whisper) or deepgram
ECHOLOOP_WHISPER_MODEL	base.en	Whisper model size
ECHOLOOP_WHISPER_DEVICE	cpu	cpu or cuda for GPU transcription
ECHOLOOP_PUSH_INTERVAL	35	Seconds between LLM coaching calls
ECHOLOOP_SILENCE_TRIGGER	4.0	Seconds of silence before early push
ECHOLOOP_SYSTEM_DEVICE		System audio device name substring
ECHOLOOP_MIC_DEVICE		Microphone device (None = default)
ECHOLOOP_MEETING_CONTEXT		One-line meeting briefing for sharper advice
ECHOLOOP_LOG_DIR		Directory for session logs (empty = disabled)
ECHOLOOP_OPACITY	0.88	Overlay window opacity

Features

• Dual audio capture: System audio + microphone, labelled [THEM] and [ME]
• Transcription: Local faster-whisper (CPU/CUDA) or Deepgram cloud
• LLM coaching: Claude or GPT every ~35s with nudge button for instant advice
• Silence detection: Skips dead air, saves CPU and API calls
• Session logging: Transcript + insights saved to disk, markdown export
• Auto-reconnect: Audio streams recover if device disconnects
• Pause/Resume: Click or Ctrl+Shift+E global hotkey
• Live stats: Insight count, session duration, talk-time percentage

Integration with Other Skills

Skill	Relationship
notebooklm	Feed meeting transcripts as sources, generate audio overviews
report-builder	Generate meeting reports from session logs
perplexity-research	Research topics mentioned in meetings in real-time
codex-companion	Delegate technical questions from meetings to Codex

Attribution

EchoLoop by Pete Woodbridge. MIT License.