Vapi Webhook Server

A webhook server that handles Vapi AI phone calls for commercial real estate lead collection. The assistant answers calls, has natural conversations with prospects, collects their information, and saves everything to a database.

What it does

• Answers inbound calls instantly with a natural-sounding AI voice
• Has conversational interactions to understand what callers are looking for
• Gently qualifies leads (property type, location, budget, timeline, etc.)
• Collects contact information (name, phone, email)
• Saves all data to SQLite database in real-time
• Stores complete call transcripts and AI-generated summaries
• Automatically logs call data to Google Sheets (optional)

Setup

• Vapi account with API key
• A publicly accessible URL for webhooks (use ngrok for testing)

Installation

uv sync

Configuration

Create a .env file:

VAPI_API_KEY=your_vapi_api_key
WEBHOOK_URL=https://your-domain.com/webhook
WEBHOOK_SECRET=your_webhook_secret
BROKERAGE_NAME=Realflow
PORT=8000

Create the assistant

python create_assistant.py

This creates a Vapi assistant configured with:

• GPT-4o for conversation handling
• Cartesia Sonic voice (natural sounding)
• Custom prompts for real estate lead qualification
• Webhook integration for data collection

Save the Assistant ID that gets printed.

Google Sheets Integration (Optional)

To automatically log all call data to a Google Sheet:

1. Follow the setup guide in GOOGLE_SHEETS_SETUP.md
2. Add your webhook URL to .env:

   GOOGLE_SHEETS_WEBHOOK_URL=https://script.google.com/macros/s/YOUR_SCRIPT_ID/exec

3. Restart your server

Call data will be logged with these columns:

• Timestamp - When the call occurred
• Name - Caller's name
• Role - owner/buyer/broker/lender
• Inquiry - What they're looking for
• Market - Location/region
• Notes - Combined notes (reason, deal size, urgency)

Running the server

python app.py

The server starts on port 8000 (or whatever you set in .env).

For local testing with ngrok:

ngrok http 8000

How it works

Call flow

1. Someone calls your Vapi phone number
2. Assistant answers immediately and introduces itself
3. Natural conversation to understand their needs
4. Collects contact details when appropriate
5. Saves information to database during the call
6. Ends call professionally with confirmation

Data storage

Everything gets saved to conversation_data/calls.db:

During the call:

• Caller information stored immediately when submitted

After the call:

• Complete transcript
• AI-generated summary
• Call duration, cost, recording URL
• Full conversation metadata

API endpoints

GET /health
GET /db/calls?limit=50

GET /db/calls/{call_id}


GET /db/stats

Database structure

The SQLite database has two main tables:

caller_information:

• Stores lead data in the exact format received from Vapi
• Includes timestamp, tool call ID, and raw payload
• Saved immediately when caller submits information

calls:

• Complete call records
• Transcripts stored as JSON
• AI summaries and success evaluations
• Call metrics and metadata

Querying the database

# Open database
sqlite3 conversation_data/calls.db

# Recent leads
SELECT * FROM caller_information ORDER BY submitted_at DESC LIMIT 10;

# Calls with summaries
SELECT call_id, summary, call_duration FROM calls ORDER BY created_at DESC;

# Export to CSV
sqlite3 -header -csv conversation_data/calls.db \
  "SELECT * FROM caller_information;" > leads.csv

Deliverables

Assistant ID: e9be887d-0a5b-485d-aba7-d6aabb556a48

I configured the webhook to my local machine and exposed it using ngrok.
The agent outputs to a database and to a webhook.