Content
# Claude Customer Support Production Capstone
A production-style AI customer support agent built using the Anthropic Claude API and MCP (Model Context Protocol).
This project evolved from a basic customer lookup application into a production-oriented support system featuring customer verification, refund processing, support ticket management, structured error handling, session management, and MCP tool integration. It also demonstrates AI-assisted development using Claude Code workflows.
---
# Features
* Customer lookup using ID, email, or name
* Order lookup
* Refund processing
* Customer verification before sensitive operations
* Support ticket creation
* Session-based customer verification
* Structured JSON error responses
* MCP Server integration
* MCP Inspector testing
* AI-assisted development with Claude Code
---
# Claude Code Tasks Completed
During the project, Claude Code was used to complete real-world engineering tasks by selecting the appropriate built-in tools (Read, Grep, Edit, Bash).
### Task 1 — Restrict Customer Data Exposure
Updated the `get_customer` tool to return only:
* customer_id
* name
* email
* account_status
instead of exposing the complete customer record.
**Result**
Improved customer privacy while preserving existing application functionality.
---
### Task 2 — Add Global Exception Handling
Implemented a catch-all exception handler in the main agent loop.
Features added:
* Structured error responses
* Exception logging
* Graceful failure handling
* Improved production reliability
---
# Project Structure
```text
claude-customer-support-production-capstone/
│
├── agent.py
├── tool_runner.py
├── tools.py
├── mock_data.py
├── mcp_server.py
├── mcp.json
├── requirements.txt
├── .env
├── .gitignore
├── README.md
│
└── screenshots/
├── 01_Task_Prompt.png
├── 02_Code_Search_Grep.png
├── 03_Proposed_Code_Change.png
├── 04_Edit_Approval.png
├── 05_Code_Modification.png
├── 06_Implementation_Progress.png
└── 07_Task_Completed.png
```
---
# Key Improvements Made
## 1. Customer Verification
Before processing refunds, customers must first verify their identity.
```python
if not session_state.get("verified_customer_id"):
return json.dumps({
"error": {
"type": "permission",
"retryable": False,
"message": "Customer must be verified before refund."
}
})
```
This prevents unauthorized refund requests.
---
## 2. Session State Management
```python
session_state = {
"verified_customer_id": None,
"verified_customer_name": None
}
```
The agent remembers the verified customer throughout the conversation.
---
## 3. Refund Processing Tool
```python
{
"name": "process_refund",
"description": "Process a refund for a verified customer."
}
```
Allows refunds only after successful verification.
---
## 4. Support Ticket Creation
```python
{
"name": "create_support_ticket",
"description": "Create a support ticket for unresolved issues."
}
```
Automatically creates tickets for unresolved customer problems.
---
## 5. Ticket Generation
```python
ticket_id = f"TICKET-{random.randint(1000,9999)}"
```
Each ticket receives a unique identifier.
---
## 6. Structured Error Handling
```python
{
"error": {
"type": "validation",
"retryable": False,
"message": "Order not found."
}
}
```
Returns standardized JSON error responses instead of generic failures.
---
## 7. MCP Tool Registration
```python
@mcp.tool()
def get_customer_tool(query: str):
return get_customer(query, session_state)
```
Exposes Python functions as MCP tools.
---
## 8. MCP Server
```python
from mcp.server.fastmcp import FastMCP
mcp = FastMCP("support-agent-tools")
```
Creates the MCP server that exposes the available tools.
---
## 9. MCP Configuration
```json
{
"mcpServers": {
"support-agent-tools": {
"command": "python",
"args": ["mcp_server.py"]
}
}
}
```
Defines how the MCP server is launched.
---
## 10. MCP Inspector Testing
Validated every tool using MCP Inspector.
Tested tools:
* get_customer_tool
* lookup_order_tool
* process_refund_tool
* create_support_ticket_tool
---
# Claude Code Workflow
This project demonstrates practical use of Claude Code for AI-assisted software development.
* Used **Grep** to locate the target function.
* Used **Read** to understand the implementation.
* Used **Edit** to apply precise code modifications.
* Approved code changes before execution.
* Verified implementation after completion.
---
# Technologies Used
* Python
* Anthropic Claude API
* MCP (Model Context Protocol)
* FastMCP
* MCP Inspector
* Claude Code
* python-dotenv
* JSON
---
# Run the Agent
```bash
python agent.py
```
---
# Run MCP Server
```bash
python mcp_server.py
```
---
# NOTE: The conversations with the claude is inside the "outputs-conversation with the claude" section
# Author
**DeepakKumar**
* Blockchain & Web3 Developer
* Smart Contract Engineer
* AI Agent Builder
