Customer Service MCP Server

Minimal FastMCP server for the DataFlow Solutions e-commerce customer-service scenario: secure read access to customers, orders, and tickets, plus write tools with audit logging and basic monitoring.

For full technical documentation (architecture, data models, security, request flows, source map), see CODEBASE.md.

What's included

Related MCP server: ShopBot AI MCP Server

Project layout

customer-service-mcp/
├── server.py           # MCP server (single file)
├── requirements.txt
├── Dockerfile
└── README.md

Setup

cd customer-service-mcp
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Configuration

Set via environment (or .env):

# key:role pairs (comma-separated)
API_KEYS=dev-admin-key:admin,dev-agent-key:agent,dev-readonly-key:readonly

LOG_LEVEL=INFO
RATE_LIMIT_PER_MIN=60
CACHE_TTL_SEC=30
MCP_TRANSPORT=stdio

Default API keys (dev only)

Run locally

python server.py

Uses stdio transport — start from Claude Desktop / Cursor MCP config, not as a standalone interactive process.

Claude Desktop example

{
  "mcpServers": {
    "customer_service": {
      "command": "/Users/amardeepganguly/building ai agents coursera/mcp-module-3/customer-service-mcp/.venv/bin/python",
      "args": [
        "/Users/amardeepganguly/building ai agents coursera/mcp-module-3/customer-service-mcp/server.py"
      ],
      "cwd": "/Users/amardeepganguly/building ai agents coursera/mcp-module-3/customer-service-mcp",
      "env": {
        "API_KEYS": "dev-admin-key:admin,dev-agent-key:agent,dev-readonly-key:readonly"
      }
    }
  }
}

Restart the host app after editing config.

Post–Claude Desktop integration test cases

After adding the server to Claude Desktop, fully quit (Cmd+Q) and reopen. Confirm customer_service shows as connected and lists the tools below.

Pre-flight

Test 1 — Read customers (agent key)

Prompt:

Use the customer_service MCP tool get_customers with api_key dev-agent-key.
Show me the customer list.

Expected: 3 customers (cust-001 Jane Johnson, cust-002 Acme Corp, cust-003 Sam Lee). No error field.

Test 2 — Orders for one customer

Prompt:

Call get_orders with api_key dev-agent-key and customer_id cust-001.

Expected: 2 orders (ord-1001, ord-1002). Statuses shipped and processing.

Test 3 — Create support ticket (write)

Prompt:

Use create_support_ticket with:
- api_key: dev-agent-key
- customer_id: cust-001
- subject: "Test ticket from Claude"
- priority: high

Expected: New ticket with id like tkt-xxxxxx, status: open. Then run Test 4 to confirm it appears.

Test 4 — Verify ticket in list

Prompt:

Call get_support_tickets with api_key dev-agent-key and customer_id cust-001.

Expected: Includes the ticket from Test 3 plus the original tkt-501.

Test 5 — Update order status (write + rollback path)

Prompt:

Call update_order_status with api_key dev-agent-key, order_id ord-1002, status shipped.
Then get_orders for cust-001 and confirm ord-1002 is shipped.

Expected: update_order_status returns order with status: shipped. get_orders reflects the change.

Test 6 — Health check (any authenticated role)

Prompt:

Run health_check with api_key dev-readonly-key.

Expected: status: healthy, circuit_open: false, timestamp present.

Test 7 — Metrics (admin only)

Prompt:

Call get_metrics with api_key dev-admin-key.

Expected: Counters include resource_reads, tool_calls, started_at. Values should be > 0 after Tests 1–6.

Test 8 — RBAC: readonly cannot write

Prompt:

Try create_support_ticket with api_key dev-readonly-key, customer_id cust-001, subject "Should fail", priority low.

Expected: error mentioning role cannot perform tickets:write (or permission denied). No new ticket created.

Test 9 — RBAC: agent cannot read metrics

Prompt:

Call get_metrics with api_key dev-agent-key.

Expected: error — role cannot perform metrics:read.

Test 10 — Invalid API key

Prompt:

Call get_customers with api_key bad-key-123.

Expected: error: invalid or missing API key. Check log: auth_failures increments (visible via Test 7 after using admin key).

Test 11 — Validation error (safe message, no stack trace)

Prompt:

Call update_order_status with api_key dev-agent-key, order_id ord-9999, status shipped.

Expected: error: unknown order_id. No internal exception text leaked to Claude.

Audit log spot-check

While running tests, watch:

tail -f ~/Library/Logs/Claude/mcp-server-customer_service.log

You should see AUDIT JSON lines for reads and writes, e.g. resource.read, tool.create_support_ticket, tool.update_order_status.

Quick smoke prompt (all-in-one)

Using customer_service MCP and api_key dev-agent-key:
1) get_customers
2) get_orders for cust-001
3) create_support_ticket for cust-001 — subject "Claude integration test" priority medium
4) health_check with dev-readonly-key
5) get_metrics with dev-admin-key
Summarize results and flag any errors.

Example prompts

Use customer_service MCP with api_key dev-agent-key:
- get_orders for customer cust-001
- create_support_ticket for cust-001 subject "Where is my order?" priority high

health_check with dev-readonly-key
get_metrics with dev-admin-key

Docker

docker build -t customer-service-mcp .
docker run --rm -i \
  -e API_KEYS=dev-admin-key:admin,dev-agent-key:agent,dev-readonly-key:readonly \
  customer-service-mcp

For production, terminate TLS at a reverse proxy if you expose HTTP/SSE transport; stdio MCP inside the container stays unencrypted (host manages the pipe).

Security notes

• Never commit real API keys.

• Errors returned to the client are sanitized (no stack traces).

• Tool inputs are validated and length-limited.

• Audit logs go to stderr — ship to your log aggregator in prod.

• Rotate keys and use distinct keys per role/tenant.

Assignment mapping

Covers the course checklist in a minimal form: server class, API key auth, RBAC, 3 resource types, 2 write tools, audit logging, health/metrics, rate limit, cache, circuit breaker, graceful errors, Docker.