---

name: teams-meeting-audio description: Join web meetings (Teams, Zoom, Google Meet, etc.) with real-time audio. AI speaks via TTS, listens via STT, and can perform web searches and meeting actions via voice commands.

Meeting Audio Bridge

Real-time AI participation in browser-based video meetings using headless Chrome, PulseAudio virtual devices, and OpenAI Realtime API.

Personalization

The default wake word and persona is "Claw". To match your agent's identity:

1. Replace "Claw" with the name from your SOUL.md (e.g., if your agent is "Alice", users will say "Alice, look up X").
2. Set the SYSTEM_PROMPT env var to override the default prompt with your agent's personality and name.
3. Alternatively, edit the SYSTEM_PROMPT constant in realtime-hybrid.js directly.

The bridge's default prompt, tool descriptions, and transcript labels all use "Claw" as a placeholder. Swap it for your agent's name everywhere it appears.

Platform Support

Works with any browser-based meeting platform:

• Microsoft Teams — Web app (teams.microsoft.com)
• Zoom — Browser client (zoom.us)
• Google Meet — Browser (meet.google.com)
• WebEx, GoToMeeting, etc. — Any web-based platform

Architecture

┌─────────────────┐     ┌──────────────────┐     ┌─────────────────┐
│   Meeting App   │────▶│ meeting-output   │────▶│ OpenAI Realtime │
│   (Chrome)      │     │   (capture)      │     │   (STT + VAD)   │
└─────────────────┘     └──────────────────┘     └────────┬────────┘
        ▲                                                  │
        │                                                  ▼
┌───────┴─────────┐     ┌──────────────────┐     ┌─────────────────┐
│   VirtualMic    │◀────│   ElevenLabs     │◀────│   GPT-4o-mini   │
│   (playback)    │     │     (TTS)        │     │   (response)    │
└─────────────────┘     └──────────────────┘     └─────────────────┘

Requirements

System

• Ubuntu VPS (headless supported)
• PulseAudio (sudo apt install pulseaudio pulseaudio-utils)
• Node.js 18+ (npm install ws in skill directory)
• Chrome with audio flags (see chrome-audio wrapper script)

Environment Variables (REQUIRED)

Set these before running:

export OPENAI_API_KEY="sk-..."
export ELEVENLABS_API_KEY="..."
export EXA_API_KEY="..."              # Optional, for web search
export ELEVENLABS_VOICE_ID="..."      # Optional, defaults to Jessica
export SYSTEM_PROMPT="..."            # Optional, override agent persona
export SAMPLE_RATE="24000"            # Optional, default 24000
export INPUT_DEVICE="meeting-output.monitor"   # Optional
export OUTPUT_DEVICE="VirtualMic"     # Optional

DO NOT commit these to the repo. Use .env files or your shell profile.

Quick Start

1. Initial Setup (One-time)

# Create PulseAudio devices (run this after reboots)
./pa-setup.sh

# Create Chrome wrapper script (save to ~/.local/bin/chrome-audio)
# See "Chrome Wrapper Script" section below

# Configure OpenClaw to use the audio-enabled Chrome
openclaw config set browser.executablePath "$HOME/.local/bin/chrome-audio"

2. Chrome Wrapper Script

Create ~/.local/bin/chrome-audio:

#!/bin/bash
exec /usr/bin/google-chrome \
  --headless=new \
  --no-sandbox \
  --disable-gpu \
  --disable-dev-shm-usage \
  --autoplay-policy=no-user-gesture-required \
  --disable-features=AudioServiceOutOfProcess \
  --use-fake-ui-for-media-stream \
  --alsa-input-device=pulse \
  --alsa-output-device=pulse \
  --user-data-dir=/tmp/chrome-meeting \
  --remote-debugging-port=18800 \
  "${@:-about:blank}"

Make executable: chmod +x ~/.local/bin/chrome-audio

Why this is critical: The --headless=new + PulseAudio flags allow Chrome to access virtual audio devices in a headless environment.

3. Join a Meeting

Option A: Automated (Recommended)

./join.sh https://meet.google.com/abc-defg-hij

This script:

• Verifies env vars
• Sets up PulseAudio devices
• Launches Chrome with the meeting URL
• Runs pre-flight audio checks
• Starts the bridge

Option B: Manual (for debugging)

# Step 1: Ensure PulseAudio is ready
./pa-setup.sh

# Step 2: Open meeting via OpenClaw browser tool
# (Browser control handles navigation, auto-join logic)
browser open --url "https://teams.microsoft.com/meet/..."

# Step 3: Start the bridge
node realtime-hybrid.js

4. In-Meeting Audio Routing

CRITICAL: Configure meeting device settings:

| Setting | Value | Why | |---------|-------|-----| | Microphone | Virtual_Microphone | Receives TTS output | | Speaker | MeetingOutput | Captures incoming audio for STT | | Camera | Off / None | We don't send video |

In Teams: Settings → Devices → Set Speaker to "MeetingOutput" In Google Meet: Settings → Audio → Speaker → "MeetingOutput" In Zoom: Settings → Audio → Speaker → "MeetingOutput"

Voice Commands

The AI responds to its name followed by requests. Replace "Claw" with your agent's name from SOUL.md.

| Command | Action | Response | |---------|--------|----------| | "Claw, what's the weather?" | Query | Spoken answer | | "Claw, search for..." | Web search (Exa) | Search results spoken | | "Claw, note that..." | Note-taking | Acknowledges and echoes to transcript | | "Claw, schedule a follow-up" | Deferred task | Accepts, logged for agent pickup | | "Claw, send the summary" | Deferred task | Accepts, logged for agent pickup | | "Claw, leave the meeting" | Exit call | "Okay, signing off!" → Leaves & stops bridge | | "Claw, mute yourself" | Stop speaking | Silences TTS (still listens) | | "Claw, unmute" | Resume speaking | Re-enables TTS | | "Claw, stop listening" | Pause STT | Stops transcription | | "Claw, start listening again" | Resume STT | Re-enables transcription |

Deferred Tasks (Transcript-Driven)

The bridge does NOT have live access to calendars, email, or other tools. Instead:

1. User asks: "Claw, schedule a meeting for Tuesday"
2. Bridge responds: "I'll set that up after the call."
3. The request appears in the transcript (📝 User: / 🗣️ Claw:)
4. After the meeting, the parent agent (OpenClaw) reviews the transcript and executes deferred tasks

This applies to: scheduling, email, reminders, calendar checks, file sharing, and any task requiring external tools.

The bridge never refuses these requests — it accepts them and defers to the agent.

Scripts Reference

pa-setup.sh — Audio Device Setup

Creates PulseAudio virtual sinks/sources at 24kHz for compatibility with OpenAI/ElevenLabs.

# Run after system reboot or if devices are missing
./pa-setup.sh

join.sh — Automated Meeting Entry

One-command launcher that orchestrates the entire pipeline.

./join.sh <meeting-url>

# Skip audio setup (use if already configured)
./join.sh --quick <meeting-url>

realtime-hybrid.js — Main Bridge

The AI brain. Handles:

• OpenAI Realtime API for speech detection
• GPT-4o-mini for response generation
• Exa search integration
• ElevenLabs TTS
• Voice command parsing

# Environment variables required
node realtime-hybrid.js

Troubleshooting

"No microphone detected" in meeting

Chrome can't see the virtual devices. Solutions:

# 1. Verify devices exist
pactl list sources short | grep Virtual

# 2. Restart PulseAudio
./pa-setup.sh

# 3. Restart Chrome (must launch AFTER devices exist)
pkill -f chrome-audio
./join.sh <url>  # Or use browser tool

Audio flowing but no transcription

# Verify audio is actually reaching the sink
timeout 2 parec --device=meeting-output.monitor | wc -c
# Should show >0 bytes

# Check bridge logs for STT errors
process log --sessionId <id>

TTS not playing in meeting

# Verify VirtualMic isn't suspended
pactl list sinks short
# Should show VirtualMic as IDLE or RUNNING, not SUSPENDED

# Test TTS manually
tts "Can you hear me?"
paplay --device=VirtualMic <tts-output-file>

Exa search not working

# Verify key is set
echo $EXA_API_KEY

# Check bridge output for "EXA_API_KEY not set"

Meeting Transcripts

After the meeting, retrieve the conversation:

# If bridge is still running
process log --sessionId <session-id> | grep -E "^📝|^🗣️"

# Or if you saved output
cat bridge.log | grep -E "^📝|^🗣️"

Format:

• 📝 User: "..." — What was said to the meeting
• 🗣️ Claw: "..." — AI responses
• 🔧 Tool call: ... — Actions taken (search, leave, etc.)

Chrome Audio Flags Explained

| Flag | Purpose | |------|---------| | --use-fake-ui-for-media-stream | Auto-grant mic/camera permissions without GUI | | --autoplay-policy=no-user-gesture-required | Allow audio playback without user click | | --disable-features=AudioServiceOutOfProcess | Keep audio in main process (headless reliability) | | --alsa-input-device=pulse | Route mic through PulseAudio | | --alsa-output-device=pulse | Route speaker through PulseAudio | | --headless=new | Modern headless mode with full web audio support |

Security Notes

• Never commit API keys to this repo. Use env vars or .env files ignored by git.
• The bridge emits SIGNAL:* to stdout for meeting controls. The parent agent (OpenClaw) must poll and execute these.
• Chrome runs in --no-sandbox mode for headless VPS compatibility. Acceptable for isolated meeting use, but don't browse untrusted sites.

Development

To modify behavior:

1. Response personality: Set SYSTEM_PROMPT env var or edit the constant in realtime-hybrid.js. Replace "Claw" with your agent's name from SOUL.md.
2. Voice: Change ELEVENLABS_VOICE_ID env var or default in code.
3. Search provider: Currently Exa. To switch, replace searchExa() function.
4. New commands: Add to MEETING_TOOLS array and handleToolCall() function.

Latency

Typical round-trip: 2-3 seconds

| Component | Time | |-----------|------| | VAD detection | ~300ms | | OpenAI transcription | ~500ms | | GPT-4o-mini response | ~800ms | | ElevenLabs TTS | ~600ms | | Audio playback | ~200ms |

Related

• PulseAudio docs: https://www.freedesktop.org/wiki/Software/PulseAudio/
• OpenAI Realtime API: https://platform.openai.com/docs/guides/realtime
• ElevenLabs TTS: https://elevenlabs.io/docs