RK
RefundKit
GitHubGet Started

OpenAI Agents SDK Integration

OpenAI Agents SDK Integration

The OpenAI Agents SDK supports MCP servers as tool providers. This guide shows how to connect the RefundKit MCP server to an OpenAI agent so it can process refunds during conversations.

Prerequisites

  • Python 3.10+
  • openai-agents package installed
  • Node.js 18+ (for the MCP server)
  • A RefundKit API key

Installation

pip install openai-agents
npm install @refundkit/sdk

Basic Setup

The OpenAI Agents SDK can connect to MCP servers using stdio transport:

import asyncio
import os
from agents import Agent, Runner
from agents.mcp import MCPServerStdio

async def main():
    # Configure the RefundKit MCP server
    refundkit_server = MCPServerStdio(
        name="refundkit",
        command="npx",
        args=["@refundkit/sdk"],
        env={
            "REFUNDKIT_API_KEY": os.environ["REFUNDKIT_API_KEY"],
        },
    )

    # Create an agent with access to RefundKit tools
    agent = Agent(
        name="Customer Support",
        instructions="""You are a customer support agent. You can process
        refunds using RefundKit tools. Always check the refund policy
        before processing a refund. Be clear about refund amounts
        and timelines.""",
        mcp_servers=[refundkit_server],
    )

    # Run the agent
    async with refundkit_server:
        result = await Runner.run(
            agent,
            "Customer wants a refund for transaction ch_1N4HbSKz9cXRvFYr. "
            "The product was defective and the charge was $25.00.",
        )
        print(result.final_output)

asyncio.run(main())

Multi-Agent Setup

You can create specialized agents that share access to RefundKit tools:

import asyncio
import os
from agents import Agent, Runner
from agents.mcp import MCPServerStdio

async def main():
    refundkit_server = MCPServerStdio(
        name="refundkit",
        command="npx",
        args=["@refundkit/sdk"],
        env={"REFUNDKIT_API_KEY": os.environ["REFUNDKIT_API_KEY"]},
    )

    # Policy checker agent
    policy_agent = Agent(
        name="Policy Checker",
        instructions="""Check refund eligibility using refundkit_get_policy.
        Report whether the refund is eligible and any conditions.""",
        mcp_servers=[refundkit_server],
    )

    # Refund processor agent
    refund_agent = Agent(
        name="Refund Processor",
        instructions="""Process approved refunds using refundkit_process_refund.
        Always confirm the amount and reason before processing.""",
        mcp_servers=[refundkit_server],
    )

    # Orchestrator agent
    orchestrator = Agent(
        name="Support Lead",
        instructions="""You coordinate refund requests. First delegate to
        the Policy Checker to verify eligibility, then delegate to
        the Refund Processor to execute the refund.""",
        handoffs=[policy_agent, refund_agent],
    )

    async with refundkit_server:
        result = await Runner.run(
            orchestrator,
            "Process a $50 refund for transaction ch_2M3GaSJy8bWqEXQp. "
            "Reason: wrong product shipped.",
        )
        print(result.final_output)

asyncio.run(main())

HTTP Transport

For production deployments, you can connect to a remote RefundKit MCP server over HTTP:

from agents.mcp import MCPServerHTTP

refundkit_server = MCPServerHTTP(
    name="refundkit",
    url="https://your-mcp-server.example.com/sse",
)

Error Handling

The OpenAI Agents SDK handles MCP tool errors automatically. When a RefundKit tool returns an error (e.g., invalid transaction ID), the agent receives the error message and can communicate it to the user naturally.

For programmatic error handling, wrap the runner call:

from agents import RunError

try:
    result = await Runner.run(agent, user_message)
except RunError as e:
    print(f"Agent run failed: {e}")

Next Steps