CainAble

Two-machine presenting from one Chrome session.

• Cain (presenter laptop): Chrome extension. Mirrors the current tab, captures call audio (Meet / Zoom web) and your mic, and sends the focused text box on a hotkey.
• Able (laptop on the big monitor): a web page served by the backend. Shows the mirrored tab, live captions, and Claude's streamed answers. The audience can ask questions here too. Answers appear only on Able.

Cain extension ──WebRTC video (via TURN)──▶ Able page
      │  room WS (signaling, text box)          ▲  room WS (transcripts, answers)
      │  audio WS (16 kHz PCM, 2ch)              │
      └────────────▶  FastAPI backend  ──────────┘
                        ├─ Deepgram Nova-3 streaming STT (ch0 = call, ch1 = presenter mic)
                        └─ Claude Opus 5.5 (streaming, effort=low, screen frame attached)

Install

pip install cainable        # or: uv tool install cainable / pipx install cainable

Requires Python 3.11+ and Chrome 116+.

Setup

1. Backend

mkdir cainable-run && cd cainable-run
cainable init               # writes .env; fill in ANTHROPIC_API_KEY, DEEPGRAM_API_KEY, TURN creds, ACCESS_KEY
cainable serve              # http://0.0.0.0:7444  (--port, --host, --env-file to change)

Both machines must reach it over HTTPS. Either deploy it (Fly.io, Render, a VM), or tunnel it: cloudflared tunnel --url http://localhost:7444 gives you a public https://….trycloudflare.com URL.

TURN is needed over the internet. Without it, most home and corporate NATs block the video. The simplest option is Cloudflare Realtime TURN (generous free tier): create a TURN key in the Cloudflare dashboard and put the key ID and API token in .env. The backend hands out short-lived credentials.

2. Able (display machine)

Open https://<backend>/?key=<ACCESS_KEY> and press F for full screen. Set ACCESS_KEY in .env; if you leave it empty, the backend generates one and prints it at startup. It shows a 6-digit room code. The code stays in the URL hash, so a reload rejoins the same room.

3. Cain (presenter machine)

1. Run cainable extension (copies the extension to ./cainable-extension), then chrome://extensions → Developer mode → Load unpacked → select that folder.
2. Open the popup, enter the backend URL, room code and access key, then Connect.
3. Go to the tab you want to show and press Alt+Shift+M (or Mirror this tab). Press it again on a different tab to switch. The video track is swapped with no reconnect.
4. Optional:
   • Alt+Shift+A on the Meet / Zoom web tab marks it as call audio. Without it, the mirrored tab's audio is transcribed.
   • My mic → Turn on: the first time, it opens a tab to grant mic permission.
   • Alt+Shift+S sends the focused text box (or the selected text) to Able, which answers it.

Rebind shortcuts at chrome://extensions/shortcuts.

How answers are triggered (all from Able)

Each question carries the current mirrored frame (toggle in ⚙), the rolling transcript, the latest shared text, and the last 6 Q&A turns.

Latency and quality knobs

• CLAUDE_EFFORT=low: fastest first token on Opus 5.5. Raise it for harder questions.
• CLAUDE_FAST_MODE=1: Opus fast mode, roughly 2.5x faster output at 2x the price (Claude API only).
• Server-side refusal fallbacks (fallbacks="default") are enabled, so a declined request is retried on a fallback model.
• Video: VP9 preferred, 6 Mbps cap, contentHint=detail, maintain-resolution. Slide text stays sharp at the cost of frame rate on bad links.

Development

git clone <repo> && cd CainAble
uv sync                      # installs the package in editable mode plus dev tools
uv run pytest
uv run cainable serve --reload

Layout: src/cainable/ holds the backend (app.py, stt.py, llm.py, ice.py, rooms.py), the Able page (able/) and the Chrome extension (extension/).

Limitations

• Call audio works only for browser calls. Zoom / Teams desktop apps need a virtual audio device (e.g. BlackHole on macOS) selected as the mic.
• The text grab can't read canvas-rendered editors (Google Docs, Figma). Select the text instead, or use Able's ask box.
• Switching the mirrored tab needs the hotkey or popup. Chrome requires a user gesture per tab capture.
• Rooms live in memory. Restarting the backend creates new codes (Able picks a new one automatically).