MCP Server Template for OpenAI Agent Builder

A minimal, reusable Model Context Protocol (MCP) server template designed for use with OpenAI Agent Builder and other MCP-compatible AI platforms. This template uses Streamable HTTP transport for broad compatibility.

Design Philosophy

This template is intentionally minimal and UI-agnostic. It provides only the MCP server infrastructure, allowing you to:

• Connect it to any AI interface (OpenAI Agent Builder, Claude, custom UIs, etc.)
• Build your own frontend or use existing AI chat interfaces.
• Focus on your business logic without UI constraints.
• Deploy as a standalone microservice.

The template includes a working example (Lebanon, NH zoning lookup) that demonstrates the pattern. Replace it with your own data source and tools.

Quick Start

npm install
npm start

The server runs on port 5000. The MCP endpoint is available at POST /mcp.

Customization Checklist

When forking this template, update the following:

Required Changes

• config.js - Update all values:

  • SERVER_NAME - Your project name (e.g., "anytown-permits")
  • LOCATION_NAME - Your municipality/region (e.g., "Anytown, USA")
  • DATA_SOURCES - Your data source names
  • TOOL_DESCRIPTIONS - Descriptions for your tools
  • ERROR_EXAMPLES - Example values for your location
  • ZONING_LAYER, ADDRESS_LAYER - Your layer IDs

• Secrets - Add required secrets:

  • ARCGIS_BASE_URL - Your FeatureServer endpoint URL (see Environment Variables section)

• package.json - Update:

  • name - Your package name
  • description - What your server does
  • author - Your name or organization
  • keywords - Relevant keywords

• mcp-server.js - Update business logic:

  • Modify lookupZoningByCoordinates() and lookupZoningByAddress() for your data
  • Update TOOLS array with your tool definitions
  • Update TOOL_HANDLERS to map tools to functions

Optional Changes

• .env - Copy from .env.example and set your environment variables
• README.md - Update documentation for your use case

Using with OpenAI Agent Builder

1. Deploy this server to a publicly accessible URL (e.g., using Replit Deployments)
2. In OpenAI Agent Builder, add a new MCP tool
3. Enter your server URL with the /mcp endpoint (e.g., https://your-app.replit.app/mcp)
4. The Agent Builder will automatically discover your available tools

Project Structure

config.js          # All customizable configuration (START HERE)
mcp-server.js      # Main server file with business logic
package.json       # Dependencies and scripts
.env.example       # Example environment variables
.gitignore         # Git ignore rules

Customizing for Your Use Case

Step 1: Update Configuration (config.js)

All location-specific strings and API endpoints are centralized in config.js:

export const SERVER_NAME = "my-city-lookup";
export const LOCATION_NAME = "My City, ST";
export const DATA_SOURCES = {
  zoningLayer: "My City Zoning Data",
  addressTable: "My City Address Database",
};

Step 2: Update Business Logic (mcp-server.js)

The server is organized into sections:

Section 1: Business Logic Functions Implement your tool functions here. Each function should:

• Accept the parameters defined in your tool's inputSchema
• Return a result object (will be JSON stringified for the AI)
• Throw an Error with a helpful message if something goes wrong

Section 2: Tool Definitions Define your MCP tools with:

• name: Unique identifier (snake_case recommended)
• description: What the tool does (imported from config.js)
• inputSchema: JSON Schema defining the parameters

Then add handlers in TOOL_HANDLERS to connect tools to your functions.

Section 3: MCP Boilerplate Standard MCP protocol handling. Rarely needs modification.

Example Tool Definition

const TOOLS = [
  {
    name: "my_custom_tool",
    description: "Description of what this tool does",
    inputSchema: {
      type: "object",
      properties: {
        param1: {
          type: "string",
          description: "Description of param1",
        },
      },
      required: ["param1"],
    },
  },
];

const TOOL_HANDLERS = {
  my_custom_tool: async (args) => {
    return await myFunction(args.param1);
  },
};

API Endpoints

Endpoint	Method	Description
/mcp	POST	Main MCP endpoint (initialize, tools/list, tools/call)
/mcp	GET	Server info (without session) or SSE stream (with session)
/mcp	DELETE	Session termination
/health	GET	Health check

Example Response

When an AI calls your tool, it receives a JSON response:

{
  "found": true,
  "district": "R-1",
  "attributes": { ... },
  "source": "Your Data Source"
}

Environment Variables & Secrets

Required Secret

This server requires the data source URL to be stored as a secret (not in code):

Secret Name	Description
ARCGIS_BASE_URL	Your FeatureServer endpoint URL (required)

On Replit: Add this in the Secrets tab (padlock icon in the Tools panel).

Why a secret? While the endpoint may be publicly accessible, storing it as a secret keeps it out of the public codebase and prevents unintended discovery.

Optional Environment Variables

These can be set as regular environment variables or secrets:

PORT=5000                    # Server port (default: 5000)
ZONING_LAYER=24              # Your zoning layer ID
ADDRESS_LAYER=6              # Your address layer ID

Features

• Streamable HTTP Transport: Works with OpenAI Agent Builder and MCP Inspector
• Stateless Client Support: Auto-initializes sessions for clients that don't handle the MCP handshake
• CORS Enabled: Ready for cross-origin requests
• Health Check Endpoint: For monitoring and load balancers
• Centralized Configuration: Easy customization via config.js

Dependencies

• @modelcontextprotocol/sdk: MCP protocol implementation
• express: HTTP server
• cors: Cross-origin support

License

ISC License - See LICENSE file for details.