galian for Cursuri AI

Posted on Apr 19

MCP (Model Context Protocol): The Complete Guide to Building AI-Powered Integrations in 2026

#programming #mcp #ai #webdev

Every developer building AI apps hits the same problem: connecting an LLM to real tools means writing custom glue code for every single integration. Different schemas, different auth, different error handling — repeated for every model and every data source.

MCP (Model Context Protocol) fixes this. It's an open standard — think USB-C for AI connectivity — that lets any AI client talk to any tool server through one universal interface. And it's not theoretical: OpenAI, Google, Microsoft, Salesforce, and thousands of developers already use it in production.

What MCP Actually Does

Before MCP, connecting Claude or GPT to your database meant writing a custom function, defining a JSON schema, handling auth, and repeating all of that for every tool. Scale that to 30 integrations across multiple environments — it breaks fast.

MCP replaces all of that with a single protocol based on JSON-RPC 2.0. A server declares what it can do; a client discovers it automatically. No hardcoding.

  Your App (Host)  →  MCP Client  →  MCP Server (tools, data, prompts)

A server can expose three things:

Tools — functions the AI can call (query_database, send_email)
Resources — structured data it can read (schemas, file contents)
Prompts — reusable templates (code review checklist, SQL generator)

A Working Example in Python

from fastmcp import FastMCP

mcp = FastMCP("Database Assistant")

@mcp.tool()
async def query_users(status: str = "active") -> list[dict]:
    """Query users filtered by status."""
    async with get_db_connection() as conn:
        rows = await conn.fetch(
            "SELECT id, name, email FROM users WHERE status = $1", status
        )
        return [dict(row) for row in rows]

@mcp.resource("schema://users")
async def get_users_schema() -> str:
    """Returns the users table schema."""
    return "CREATE TABLE users (id SERIAL PRIMARY KEY, name VARCHAR, email VARCHAR, status VARCHAR);"

mcp.run(transport="stdio")

15 lines. Your AI agent can now query your database and understand its schema through any MCP-compatible client.

TypeScript Works Too

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

const server = new McpServer({ name: "GitHub Assistant", version: "1.0.0" });

server.tool(
  "list_issues",
  "List open issues for a repository",
  { owner: z.string(), repo: z.string(), limit: z.number().default(10) },
  async ({ owner, repo, limit }) => {
    const res = await fetch(
      `https://api.github.com/repos/${owner}/${repo}/issues?state=open&per_page=${limit}`
    );
    return { content: [{ type: "text", text: JSON.stringify(await res.json(), null, 2) }] };
  }
);

await server.connect(new StdioServerTransport());

Two Transports, Different Use Cases

stdio — local tools. Server runs as a child process, zero network overhead. Great for file access, local DBs, CLI tools.

Streamable HTTP — remote/shared servers. Runs as a web service, supports OAuth 2.0. Ideal for SaaS integrations and team-shared tools.

Most production setups use both.

Why MCP Won

The adoption timeline tells the story:

Nov 2024 — Anthropic launches MCP as open-source
Mar 2025 — OpenAI adopts MCP officially
May 2025 — Microsoft joins the MCP steering committee
Jun 2025 — Salesforce builds Agentforce 3 on MCP
Dec 2025 — MCP moves to the Linux Foundation

Today: 10,000+ servers in production, 70%+ of major SaaS brands ship MCP servers, every major AI platform supports it.

Security Done Right

MCP's security model is one of its strongest features:

Granular permissions — each server declares capabilities, the host controls access
User consent — critical actions need explicit approval
Process isolation — servers run in separate processes
Full audit trail — every invocation is logged

From Demo to Production

A tutorial MCP server and a production one are very different. Production needs OAuth 2.0, rate limiting, Docker/Kubernetes deployment, CI/CD pipelines, GDPR compliance, and threat modeling.

If you want the full path — from fundamentals to deploying enterprise-grade MCP servers with Python and TypeScript — check out this complete MCP course. 24 hours of hands-on content with real projects: PostgreSQL, external APIs, multi-server gateways, and production security patterns.

Start Here

Install Claude Desktop or Cursor as your MCP host
Try a pre-built server (filesystem, PostgreSQL)
Build a custom server with FastMCP or the TypeScript SDK
Add HTTP transport and OAuth for remote access
Deploy with Docker

MCP is infrastructure, not a trend. The developers who learn it now will build the next generation of AI applications.

Want more production-focused AI engineering content? Visit Cursuri-AI.ro — courses built for developers who ship.

DEV Community