A minimal MCP (Model Context Protocol) server example. Use this as a starting point for building your own MCP tools.

const response = await openai.responses.create({
  model: "gpt-4o-mini",
  input: "Roll 3 dice for me",
  tools: [{
    type: "mcp",
    server_label: "dice",
    server_url: "YOUR_VAL_URL_HERE",
    require_approval: "never",
  }],
});

Add this server in the MCP Servers settings panel.

# List tools
curl -X POST YOUR_VAL_URL_HERE \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

# Roll dice
curl -X POST YOUR_VAL_URL_HERE \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"roll_dice","arguments":{"count":3}}}'

This template shows the essential MCP pattern:

1. Define tools with name, description, and JSON Schema for inputs
2. Handle protocol methods: initialize, tools/list, tools/call
3. Return JSON-RPC 2.0 responses

Key methods to implement:

• initialize - Return server capabilities
• tools/list - Return array of available tools
• tools/call - Execute a tool and return results

See the MCP specification for the full protocol.