AI Customer Support Agent

This project is a production-style starter for an AI customer support system built with FastAPI, a lightweight MCP-compatible tool server, and an OpenAI-compatible LLM client.

Features

• Answers customer questions with a knowledge base connector

• Retrieves order status from a database

• Creates support tickets for unresolved issues

• Summarizes conversations for human agents

• Detects likely escalation scenarios

• Exposes required tools through an MCP-style JSON-RPC endpoint

Related MCP server: Acme Operations Assistant

Project Structure

ai-support-agent/
├── backend/
│   ├── agent.py
│   ├── main.py
│   ├── mcp_server.py
│   ├── database/
│   │   └── models.py
│   └── tools/
│       ├── crm_tool.py
│       ├── kb_tool.py
│       ├── order_tool.py
│       └── ticket_tool.py
├── frontend/
│   └── simple_chat_ui.html
├── requirements.txt
└── README.md

Architecture

• FastAPI backend: serves the /chat API, health endpoint, and the simple browser UI.

• SupportAgent: orchestrates LLM responses and decides when to call tools.

• MCP server: exposes get_order_status, search_knowledge_base, create_support_ticket, and get_customer_details over JSON-RPC at /mcp.

• Connectors: isolated tool classes for orders, CRM, knowledge base, and ticketing.

• Database: SQLAlchemy models for customers, orders, and support_tickets.

Tech Choices

• Python 3.11+

• FastAPI

• SQLAlchemy

• SQLite by default, with a clear upgrade path to PostgreSQL

• OpenAI-compatible SDK via the openai Python package

Setup

1. Create and activate a virtual environment.

2. Install dependencies:

pip install -r requirements.txt

3. Optional: configure an OpenAI-compatible provider.

export OPENAI_API_KEY="your_api_key_here"
export OPENAI_MODEL="gpt-4.1-mini"
export OPENAI_BASE_URL="https://api.openai.com/v1"

If OPENAI_API_KEY is not set, the app still runs in a rules-based fallback mode so you can test the flows locally.

Run

uvicorn backend.main:app --reload --host 0.0.0.0 --port 8000

Then open http://localhost:8000.

API Usage

POST /chat

Request:

{
  "customer_id": 1,
  "message": "Where is my order #45231?",
  "conversation_history": []
}

Response:

{
  "response": "Order #45231 for Noise-Cancelling Headphones is currently shipped. Carrier: FedEx. Tracking: ZX991245US. Estimated delivery: 2026-03-15.",
  "used_tools": [
    {
      "name": "get_order_status",
      "arguments": {
        "order_id": "45231"
      },
      "result": {
        "order_id": 45231,
        "customer_id": 1,
        "item_name": "Noise-Cancelling Headphones",
        "status": "shipped",
        "tracking_number": "ZX991245US",
        "shipping_carrier": "FedEx",
        "estimated_delivery": "2026-03-15",
        "total_amount": 199.99
      }
    }
  ],
  "escalated": false,
  "conversation_summary": "Customer 1 asked: Where is my order #45231?. Agent responded: ...",
  "llm_mode": false
}

POST /mcp

Example initialization request:

{
  "jsonrpc": "2.0",
  "id": "1",
  "method": "initialize",
  "params": {}
}

Example tool list request:

{
  "jsonrpc": "2.0",
  "id": "2",
  "method": "tools/list",
  "params": {}
}

Example tool call request:

{
  "jsonrpc": "2.0",
  "id": "3",
  "method": "tools/call",
  "params": {
    "name": "get_order_status",
    "arguments": {
      "order_id": "45231"
    }
  }
}

Database

The app creates and seeds a local SQLite database file named support_agent.db on startup with example customers, orders, and support tickets.

To migrate to PostgreSQL:

• replace DATABASE_URL in backend/database/models.py

• update the engine configuration

• add migrations with Alembic for production deployments

Example Agent Flow

User message:

Where is my order #45231?

Expected flow:

1. Agent detects an order-tracking request.

2. Agent calls get_order_status(order_id).

3. Tool returns shipping carrier, tracking number, and estimated delivery.

4. Agent responds with a concise customer-facing answer.

Error Handling

• Invalid order IDs return 400

• Unknown customers return 400

• Unexpected backend failures return 500

• Tool errors are surfaced in structured JSON-RPC format on the MCP endpoint

Notes

• The knowledge base is intentionally simple and in-memory for easy extension.

• The ticketing and CRM connectors are implemented as modular service classes so they can be swapped with real APIs later.

• For a production deployment, add authentication, persistent conversation storage, rate limiting, and observability.