How do you create an MCP server for your website?

What MCP is

The Model Context Protocol (MCP) is an open standard that lets AI applications call tools and read data from outside systems through one interface. A server offers tools (functions the model can call), resources (data it can read) and prompts (reusable templates). The current specification is dated 2026-07-28. It made the protocol stateless, so each request stands alone, and it keeps Streamable HTTP as the transport for remote servers (MCP blog).

Do you need one?

Step 1: choose the tools

Design tools around what a user asks for, not around your database tables. A store usually needs search_products, get_product, check_availability, create_cart and get_order_status. Keep the list short: every tool description is read by the model on every request, and overlapping tools cause wrong calls.

• Name tools with a verb and a noun.
• Describe when to use the tool and what it returns, in one or two sentences.
• Constrain inputs with a JSON Schema: enums, formats, required fields.
• Return compact, structured results with identifiers the next call can use. Do not return HTML.
• Mark tools that change something and require confirmation for purchases.

Step 2: write the server

Official SDKs exist for TypeScript, Python and other languages. The example below uses the TypeScript SDK. SDK APIs changed with the 2026-07-28 specification, so check the SDK documentation for the version you install.

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

const server = new McpServer({ name: 'example-store', version: '1.0.0' });

server.registerTool(
  'search_products',
  {
    title: 'Search products',
    description: 'Search the catalog by keyword. Returns up to 10 products with id, name, price, currency and availability.',
    inputSchema: { query: z.string().min(2), inStockOnly: z.boolean().optional() },
  },
  async ({ query, inStockOnly }) => {
    const items = await catalog.search(query, { inStockOnly }); // your existing search
    return { content: [{ type: 'text', text: JSON.stringify(items.slice(0, 10)) }] };
  },
);

server.registerTool(
  'get_product',
  {
    title: 'Get product',
    description: 'Get full details for one product id: price, unit, minimum order quantity, stock and lead time.',
    inputSchema: { id: z.string() },
  },
  async ({ id }) => ({ content: [{ type: 'text', text: JSON.stringify(await catalog.get(id)) }] }),
);

Step 3: serve it over HTTP

Mount the server on a public HTTPS endpoint, commonly /mcp on your own domain or an mcp. subdomain, using the SDK’s Streamable HTTP transport. Design it to be stateless: do not depend on a session living on one machine. The older HTTP+SSE transport is deprecated.

Step 4: add authorization

Read-only catalog tools can be public. Anything tied to an account needs OAuth: the MCP server acts as a protected resource, points clients to your authorization server through published metadata, and validates the access token on every request. Give tokens narrow scopes such as orders:read, and never accept a token that was issued for a different service. The current specification tightened these rules, so read its authorization section before you ship.

Step 5: make it discoverable

• Document it on your site: the endpoint URL, the tools and how to connect.
• List it in the MCP Registry (registry.modelcontextprotocol.io), which stores metadata about public servers.
• Mention it in `llms.txt` and in your developer documentation.
• Consider a server card. A discovery file under /.well-known/ is proposed and some scanners already look for it, but the path is not final. Treat it as optional.

Step 6: test it

Use the MCP Inspector to call each tool by hand, then connect the server to an AI client and ask it to complete real tasks. Watch for the model choosing the wrong tool, sending invalid arguments or needing several calls for what should be one. Each of those is a description or schema to improve.

Security checklist

• Validate every input on the server. The model’s arguments are untrusted.
• Treat text you return from user-generated content as untrusted too. It can contain instructions aimed at the model.
• Rate-limit by token and by tool.
• Require explicit user confirmation for payments and other irreversible actions.
• Log tool calls so you can see what agents do and where they fail.

Frequently asked questions