AI assistant + computer vision overlay for Google Meet.
Built with Python Flask · MediaPipe · Anthropic Claude API · Chrome Extension (MV3)

meet-ai-extension/
├── backend/
│   ├── app.py          ← Flask server (AI chat + vision API + WebSocket)
│   ├── vision.py       ← MediaPipe face mesh + attention scoring
│   └── requirements.txt
├── extension/
│   ├── manifest.json   ← Chrome Extension MV3
│   ├── content.js      ← Injected into Google Meet
│   └── sidebar/
│       └── sidebar.css ← Island styles
├── test_page/
│   └── index.html      ← Standalone test environment (no Chrome needed)
└── .vscode/
    └── launch.json     ← One-click run in VSCode

cd backend
pip install -r requirements.txt

macOS/Linux: if mediapipe fails, try pip install mediapipe --pre
Windows: use pip install mediapipe-silicon on Apple Silicon

export ANTHROPIC_API_KEY="sk-ant-..."
# Windows PowerShell:
$env:ANTHROPIC_API_KEY="sk-ant-..."

Option A — VSCode (recommended)

• Open the project folder in VSCode
• Press F5 or go to Run → Start Debugging → "▶ Run Meet AI Server"

Option B — Terminal

cd backend
python app.py

Server starts at http://localhost:5000

Open http://localhost:5000 — you'll see the full simulated Google Meet environment with the AI island.

• Click 👁 to activate computer vision (requires camera permission)
• Type in the chat box to talk to the AI assistant
• Click ⬜ to toggle compact view (shrinks island, keeps camera visible)
• Drag the island by its header to reposition
• 🧠 button toggles whether vision data is sent as context to the AI

Once you're happy testing locally, load it into Chrome:

1. Go to chrome://extensions
2. Enable Developer mode (top right toggle)
3. Click Load unpacked
4. Select the extension/ folder
5. Join any Google Meet call — the island appears automatically

The extension communicates with your local Flask server (localhost:5000), so keep it running.

The attention score (0–100%) is derived from head pose:

• > 65% → green ring (looking at screen)
• 40–65% → yellow ring (slight drift)
• < 40% → red ring + alert shown

This keeps the assistant subtle — the compact view with a small camera preview looks entirely natural in a meeting context, so participants won't notice you're consulting the AI.

Change the AI personality → edit the system prompt in backend/app.py
Change analysis frequency → adjust the 1100 ms timeout in test_page/index.html / extension/content.js
Adjust attention thresholds → edit _attention_from_pose() in backend/vision.py
Compact size → adjust .mai-compact in sidebar/sidebar.css