WhatsApp Booking Agent

AI-powered booking agent for WhatsApp built with AWS Bedrock AgentCore and Strands Agents.

Architecture

• Model: DeepSeek R1 (via AWS Bedrock)
• Framework: Strands Agents
• Platform: AWS Bedrock AgentCore
• Session Management: S3SessionManager (persists conversations per phone number)
• API Integration: AgentCore Gateway (wraps TempoBook API as MCP tools)
• API: TempoBook booking API

Booking Flow

1. User requests service via WhatsApp
2. Agent shows available services
3. User selects service(s)
4. Agent shows establishments and available dates
5. Agent displays time slots
6. User chooses time
7. Agent collects customer info (name, email, phone)
8. Agent creates booking and sends confirmation

Gateway Setup

AgentCore Gateway automatically wraps the TempoBook API as MCP tools for the agent.

Create Gateway

uv sync
uv run python create_gateway.py

This will:

1. Create IAM role for Gateway
2. Create AgentCore Gateway with MCP protocol
3. Add TempoBook API as a target using the OpenAPI spec
4. Output the Gateway MCP endpoint URL

Save the Gateway MCP URL for the next step.

Setup

Prerequisites

uv sync
aws configure

Configuration

1. Create S3 bucket for sessions:

aws s3 mb s3://tempobook-sessions

2. Set Gateway MCP URL as environment variable:

export GATEWAY_MCP_URL="<your-gateway-mcp-url>"

3. Configure AgentCore (use agent_with_gateway.py for Gateway integration):

uv run agentcore configure -e agent_with_gateway.py

4. Deploy:

uv run agentcore deploy

Note: If you prefer using direct API calls without Gateway, use agent.py instead of agent_with_gateway.py.

WhatsApp Integration

Invoke from your WhatsApp webhook:

import boto3
import json

client = boto3.client('bedrock-agentcore')

response = client.invoke_agent_runtime(
    agentRuntimeArn="<your-agent-arn>",
    runtimeSessionId=f"whatsapp-{phone_number}",
    payload=json.dumps({
        "phone": phone_number,
        "message": user_message
    }).encode()
)

reply = json.loads(response['payload'].read())['response']

Tools

AgentCore Gateway automatically exposes these tools via MCP from the TempoBook API:

getServices

• Lists all available services with duration and pricing
• Mapped from: GET /api/services

getEstablishmentsByServices

• Finds establishments offering selected services
• Args: comma-separated service IDs
• Mapped from: GET /api/establishments/establishments-by-services

getAvailableTimes

• Shows available time slots for specific date
• Args: date (ISO), establishment ID, service IDs, total duration (minutes)
• Mapped from: POST /api/bookings/availableTime

createBooking

• Creates the booking
• Args: customer info, service IDs, establishment, date, time, duration
• Mapped from: POST /api/bookings/

The OpenAPI specification is in tempobook-api-spec.yaml.

Environment Variables

Set in AgentCore configuration:

• GATEWAY_MCP_URL: MCP endpoint URL from Gateway setup
• S3 bucket name: tempobook-sessions (hardcoded in agent code)

Testing Locally

uv run agentcore dev

Or test directly:

uv run agentcore invoke '{"phone": "+593995604584", "message": "I want to book a service"}'

API Endpoints Used

• GET /api/services
• GET /api/establishments/establishments-by-services
• POST /api/bookings/availableTime
• POST /api/bookings/

Session Management

Each WhatsApp conversation is isolated by phone number:

• Session ID: whatsapp-{phone_number}
• Storage: S3 (auto-persisted by S3SessionManager)
• Conversation window: 20 messages (SlidingWindowConversationManager)

Notes

• DeepSeek R1 is cost-effective for this use case
• Sessions persist across message invocations
• Agent maintains conversation context automatically
• All dates use ISO 8601 format
• Phone codes default to international format (e.g., "+593")