How to Build an MCP Server for Your Business: A 30-Minute Tutorial

Step 1: Install the MCP SDK

The official SDK from Anthropic handles the protocol, schema validation, and SSE transport. Spin up a Node 20 project and pull in two packages.

mkdir my-business-mcp && cd my-business-mcp
npm init -y
npm install @modelcontextprotocol/sdk zod
npm install -D typescript @types/node tsx
npx tsc --init --target es2022 --module esnext --moduleResolution bundler

zod is used to describe input schemas in a way both humans and agents can reason about. The SDK turns the zod schema into a JSON Schema that agents fetch during discovery.

Step 2: Define Three Tools

Tools are the verbs of your business. Start with three that every business has: list what you sell, quote a price, and place an order. Each tool gets a name, a zod schema, and a handler.

import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
import { z } from 'zod'

const server = new McpServer({ name: 'acme-biz', version: '1.0.0' })

server.tool(
  'list_products',
  { category: z.string().optional(), limit: z.number().int().max(100).default(20) },
  async ({ category, limit }) => {
    const rows = await db.products.list({ category, limit })
    return { content: [{ type: 'text', text: JSON.stringify(rows) }] }
  }
)

server.tool(
  'get_pricing',
  { product_id: z.string(), quantity: z.number().int().positive().default(1) },
  async ({ product_id, quantity }) => {
    const price = await db.pricing.compute(product_id, quantity)
    return { content: [{ type: 'text', text: JSON.stringify(price) }] }
  }
)

server.tool(
  'create_order',
  {
    items: z.array(z.object({ product_id: z.string(), qty: z.number().int().positive() })),
    customer_email: z.string().email(),
    idempotency_key: z.string().uuid(),
  },
  async ({ items, customer_email, idempotency_key }) => {
    const order = await db.orders.createIdempotent({ items, customer_email, idempotency_key })
    return { content: [{ type: 'text', text: JSON.stringify(order) }] }
  }
)

Step 3: Add Resources (Static Docs the Agent Can Read)

Resources are read-only documents agents load once and reason over. Business hours, refund policy, terms of service. Expose them as MCP resources instead of making the agent scrape HTML.

server.resource(
  'business-info',
  'resource://acme/business-info',
  async () => ({
    contents: [{
      uri: 'resource://acme/business-info',
      mimeType: 'application/json',
      text: JSON.stringify({
        legal_name: 'Acme Inc.',
        hours: { mon_fri: '9:00-18:00', sat: '10:00-14:00', sun: 'closed' },
        phone: '+1-555-0100',
        address: '100 Market St, San Francisco, CA',
        refund_policy_url: 'https://acme.com/refunds',
      }),
    }],
  })
)

Step 4: Define a Prompt Workflow

Prompts are pre-built interaction templates. They guide an agent through multi-step flows — onboarding, quote-to-order, return-a-product. One prompt can orchestrate three tools in the right sequence.

server.prompt(
  'place_order_flow',
  'Guide a customer from product search through checkout',
  { customer_intent: z.string() },
  async ({ customer_intent }) => ({
    messages: [
      { role: 'user', content: { type: 'text', text:
        'Customer said: ' + customer_intent + '\n\n' +
        '1. Call list_products to find matches.\n' +
        '2. Call get_pricing on top 3 candidates.\n' +
        '3. Confirm with customer, then call create_order with a fresh UUID idempotency_key.'
      }}
    ]
  })
)

Step 5: Deploy

Wire your McpServer to an SSE transport and expose it at /api/mcp. Three hosting paths, ranked by effort.

For the Vercel path, add a single route file at app/api/mcp/route.ts that imports your server and pipes it through SSEServerTransport. Enable runtime = 'nodejs' and set a generous maxDuration for long-lived streams.

Step 6: Register in agent-card.json

A deployed MCP endpoint is useless if no agent can find it. The agent-card.json file at /.well-known/agent-card.json is the discovery hook agents like Claude, ChatGPT, and AgentHermes use to enumerate your capabilities.

{
  "protocolVersion": "0.3",
  "name": "Acme Inc.",
  "description": "Online retailer of industrial supplies",
  "url": "https://acme.com",
  "mcp_server": {
    "transport": "sse",
    "endpoint": "https://acme.com/api/mcp",
    "auth": { "type": "oauth2", "authorization_url": "https://acme.com/oauth/authorize" }
  },
  "skills": [
    { "id": "browse",  "tool": "list_products" },
    { "id": "quote",   "tool": "get_pricing" },
    { "id": "checkout","tool": "create_order" }
  ]
}

Test It Before You Ship

Run the MCP Inspector — the official debugging UI — against your deployed URL. It enumerates every tool, validates schemas, and lets you fire test calls. If the Inspector shows your three tools with green checkmarks, Claude and ChatGPT will see the same.

npx @modelcontextprotocol/inspector https://acme.com/api/mcp

Then run an AgentHermes scan at /audit against your domain. A working MCP endpoint plus a linked agent card lifts D2 API Quality and D9 Agent Experience — up to 12 points combined. That is enough to move most businesses from Bronze to Silver in a single deploy.

Not a Developer? Skip to AgentHermes /connect

This tutorial is aimed at engineers. If you run a business and do not write code, AgentHermes /connect auto-generates everything described above — tools tuned to your vertical, resources populated from your business profile, a hosted MCP endpoint, and a linked agent card. The output is the same MCP server this tutorial builds. You just skip the 30 minutes.

Continue Reading