Skip to content

sophie-voice-bridge

⭐ 1 stars

📊 Project Details

  • Primary Language: Python
  • Languages Used: Python, Dockerfile
  • License: None
  • Created: January 28, 2026
  • Last Updated: April 30, 2026

📝 About

Sophie Voice Bridge

ElevenLabs Custom LLM server that connects to Sophie AI via Clawdbot.

Architecture

WhatsApp Call → ElevenLabs (STT/TTS) → Haiku (fast) ─┬─→ Direct response
                                                      │
                                                      └─→ ask_sophie → Sophie/Opus
                                                                          (full capabilities)
  • Haiku handles fast voice responses (1-3 sentences)
  • ask_sophie tool escalates complex questions to Sophie (Opus with tools)
  • Sophie has access to calendar, email, web search, files, memory, etc.

Quick Start (Docker)

# Clone
git clone https://github.com/aaron777collins/sophie-voice-bridge.git
cd sophie-voice-bridge

# Configure
cp .env.example .env
# Edit .env with your Clawdbot gateway token

# Run
docker compose up -d

# Check health
curl http://localhost:8013/health

Configuration

Create .env file:

CLAWDBOT_GATEWAY_URL=http://localhost:18789
CLAWDBOT_GATEWAY_TOKEN=your-gateway-token

API Endpoints

Both routes work (ElevenLabs uses /chat/completions):

POST /v1/chat/completions    # OpenAI standard
POST /chat/completions       # ElevenLabs format
GET  /health                 # Health check

🎙️ ElevenLabs Setup Guide

Step 1: Create an Agent

  1. Go to ElevenLabs Agents Platform
  2. Click "Create Agent"
  3. Give it a name (e.g., "Sophie Voice")

Step 2: Configure LLM (Custom)

  1. In Agent settings, go to "LLM" section
  2. Select "Custom LLM"
  3. Configure:
  4. Server URL: https://voice.aaroncollins.info
  5. Model: haiku (or anything - it's ignored)
  6. Custom LLM extra body: ✅ Enable this
  7. Limit token usage: 5000 (recommended)

Step 3: Configure Voice

  1. Go to "Voice" section
  2. Choose a voice you like (the voice Sophie will speak with)
  3. Adjust stability/clarity as needed

Step 4: Connect WhatsApp (via Twilio)

  1. Go to "Phone Numbers" section
  2. Click "Import Phone Number"
  3. Enter your Twilio credentials:
  4. Phone Number: Your Twilio number with WhatsApp enabled
  5. Twilio SID: From Twilio console
  6. Twilio Token: From Twilio console
  7. Assign the agent to the phone number

Step 5: Test

Outbound call (you call someone): 1. In Phone Numbers, click the outbound call button 2. Select your agent 3. Enter the phone number to call 4. Click "Send Test Call"

Inbound call (someone calls you): - Call your Twilio number on WhatsApp - The agent will answer automatically

How It Works

  1. You speak on WhatsApp call
  2. ElevenLabs STT transcribes your speech to text
  3. Text sent to /chat/completions endpoint
  4. Haiku evaluates:
  5. Simple question → responds directly (fast, 1-3 sentences)
  6. Complex question → uses ask_sophie tool
  7. If ask_sophie used:
  8. Sophie (Opus) processes with full tool access
  9. Can check calendar, email, search web, access files
  10. Haiku relays response conversationally
  11. ElevenLabs TTS converts response to speech
  12. You hear Sophie's voice response

When ask_sophie is Used

Haiku automatically escalates to Sophie when you ask about: - Calendar, schedule, events, appointments - Emails, messages, notifications - Files, documents, code, projects - Research requiring web search or browsing - Complex technical or business questions - Anything requiring memory of past conversations

Simple greetings, math, and general knowledge are handled by Haiku directly.

Development

# Local setup
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

# Run locally
python bridge.py

Test Commands

# Health check
curl https://voice.aaroncollins.info/health

# Simple question (Haiku direct)
curl -X POST https://voice.aaroncollins.info/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": "What is 2+2?"}], "stream": false}'

# Complex question (triggers ask_sophie)
curl -X POST https://voice.aaroncollins.info/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": "What meetings do I have tomorrow?"}], "stream": false}'

License

MIT