decagon-cx-mcp-server

MCP server that connects AI customer support agents to Zendesk, Stripe, and Shopify.

Your agents already think. This gives them hands.

Built as a technical proof-of-concept showing how Model Context Protocol (MCP) can standardize the integration layer between AI agents and customer service backends. Instead of building custom API wrappers for each new enterprise customer, an MCP server provides a unified tool interface that any agent can discover and call.

graph TD
    A[Customer Query] -->|natural language| B[CX Agent<br>Claude + LangChain]
    B -->|MCP tool calls| C[MCP Server]
    C -->|route| D[Zendesk Tools]
    C -->|route| E[Stripe Tools]
    C -->|route| F[Shopify Tools]
    D -->|API call| G[Mock Zendesk API]
    E -->|API call| H[Mock Stripe API]
    F -->|API call| I[Mock Shopify API]
    C -.->|middleware| J[Auth / Rate Limiting]
    C -.->|metrics| K[Observability]
    C -.->|tenant context| L[Multi-tenant Isolation]
    style A fill:#90EE90
    style B fill:#90EE90
    style C fill:#90EE90
    style D fill:#90EE90
    style E fill:#90EE90
    style F fill:#90EE90
    style G fill:#90EE90
    style H fill:#90EE90
    style I fill:#90EE90
    style J fill:#FFD700
    style K fill:#FFD700
    style L fill:#FFD700

Green = live, working code. Yellow = blueprint (architecture documented, stubs present).

Watch full quality video

# Clone
git clone https://github.com/hashwnath/decagon-cx-mcp-server.git
cd decagon-cx-mcp-server

# Install
pip install -r requirements.txt

# Run tests (no API key needed)
python -m pytest tests/ -v

# Run the MCP server
python -m src.server

# Check server health
curl http://localhost:8080/health

# Run the agent demo (requires Anthropic API key)
export ANTHROPIC_API_KEY=your-key
python examples/run_agent.py

# Run all demo scenarios
python examples/sample_conversations.py

• zendesk_get_ticket - look up a ticket by ID
• zendesk_search_tickets - search by keyword or requester email
• zendesk_update_ticket_status - update ticket status (open, pending, solved, closed)
• zendesk_add_comment - add public or internal comments

• stripe_get_payment - look up payment by ID
• stripe_search_payments - find payments by customer email
• stripe_process_refund - process full or partial refunds with eligibility check
• stripe_get_subscription - look up subscription details
• stripe_cancel_subscription - cancel at period end or immediately

• shopify_get_order - look up order by ID
• shopify_search_orders - find orders by customer email
• shopify_get_tracking - get shipping tracking info with carrier and ETA

The agent handles multi-step resolutions end-to-end:

Order Tracking: "Where's my order SH-1234?"

• Agent calls shopify_get_order then shopify_get_tracking, returns tracking status and ETA

Refund Processing: "My headphones arrived damaged, I want a refund"

• Agent calls shopify_get_order to verify, stripe_search_payments to find payment, stripe_process_refund to execute, zendesk_update_ticket_status to close ticket

Subscription Cancellation: "I was charged twice, cancel my subscription"

• Agent calls stripe_search_payments to find charges, stripe_get_subscription for plan details, stripe_process_refund for duplicate, stripe_cancel_subscription

Every enterprise AI CX deployment faces the same bottleneck: connecting agents to customer backends. Each new customer means new API integrations. MCP standardizes this:

• Tool discovery: agents learn available actions at runtime
• Consistent interface: same protocol regardless of backend service
• Composable: add new backends without changing agent code
• Observable: centralized logging of all tool calls

Bilt deployed Decagon's AI agents and automated 70% of 60K monthly tickets. Ada doubled resolution rates by enabling agents to execute actions, not just answer questions. The pattern is clear: resolution requires tool access. MCP provides the standard.

Built by Hashwanth Sutharapu as a technical POC for Decagon.

Relevant experience:

• Production MCP servers handling 1K+ requests/min at MAQ Software (Microsoft Partner)
• Multi-agent orchestration: ThinkNet 8-agent system on AWS Bedrock
• Enterprise agent deployment for MCAPS teams at Microsoft
• Microsoft Agent Framework contributor (8K+ stars)
• MS CS from ASU (3.93 GPA)

github.com/hashwnath | linkedin.com/in/sutharapuhashwanth | hsuthara@asu.edu