What is Model Context Protocol (MCP)? A Developer Guide to AI Tool Servers

Historically, building autonomous AI agent loops meant maintaining an NxM integration nightmare: if you had 5 AI models (OpenAI, Gemini, Claude, Llama, DeepSeek) and 10 internal data sources (Postgres, Git, S3, Jira, Salesforce), you had to write and maintain 50 bespoke tool definitions.

With the industry-wide adoption of Model Context Protocol (MCP), this problem is solved. Developers write one MCP server per data source, and every compliant AI client can immediately discover, query, and execute tools securely.

1. The Three Core Primitives of MCP

An MCP server exposes functionality through three well-defined primitives:

2. Transport Layer: Stdio vs. Server-Sent Events (SSE)

MCP supports two primary transport mechanisms:

• - Stdio Transport (Local Process): The AI client spawns the MCP server as a local child process and communicates via standard input/output streams. This provides sub-millisecond execution latency with maximum local security.
• - SSE / HTTP Transport (Remote Servers): The client connects to a remote MCP server over HTTPS using Server-Sent Events for streaming messages. This enables centralized enterprise tool hubs shared across distributed engineering teams.

3. Code Example: Building a Postgres Query MCP Server in TypeScript

Here is a complete, minimal implementation of an MCP server that provides a safe SQL query execution tool using the official @modelcontextprotocol/sdk:

import { Server } from '@modelcontextprotocol/sdk/server/index.js';
import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
import { CallToolRequestSchema, ListToolsRequestSchema } from '@modelcontextprotocol/sdk/types.js';
import { Pool } from 'pg';

const pool = new Pool({ connectionString: process.env.DATABASE_URL });
const server = new Server({ name: 'postgres-mcp-server', version: '1.0.0' }, { capabilities: { tools: {} } });

server.setRequestHandler(ListToolsRequestSchema, async () => ({
  tools: [{
    name: 'execute_readonly_sql',
    description: 'Execute a read-only SELECT query against the analytics database',
    inputSchema: {
      type: 'object',
      properties: { query: { type: 'string', description: 'SQL SELECT query string' } },
      required: ['query']
    }
  }]
}));

server.setRequestHandler(CallToolRequestSchema, async (request) => {
  if (request.params.name === 'execute_readonly_sql') {
    const sql = String(request.params.arguments?.query);
    if (!sql.trim().toUpperCase().startsWith('SELECT')) {
      throw new Error('Security Violation: Only SELECT queries are permitted.');
    }
    const res = await pool.query(sql);
    return { content: [{ type: 'text', text: JSON.stringify(res.rows, null, 2) }] };
  }
  throw new Error('Unknown tool');
});

const transport = new StdioServerTransport();
await server.connect(transport);

4. Security Best Practices & Guardrail Isolation

Exposing filesystem and database capabilities to AI models requires strict AI tool safety guards:

• - Read-Only by Default: Restrict database database connection strings to read-only user roles to prevent destructive table mutations.
• - Path Whitelisting: Validate that file reading tools cannot traverse parent directories (preventing ../../etc/passwd vulnerabilities).
• - Human Confirmation Dialogs: For MCP tools that trigger emails, deploy code, or execute financial transactions, require explicit interactive user confirmations.

5. How MCP Supercharges Hybrid Reasoning Models

When paired with reasoning models like Gemini 3.7 Flash or Claude 3.7 Sonnet, MCP servers provide the structured grounding necessary to eliminate hallucinations.

Because MCP tools define strict JSON schemas, reasoning models can utilize dynamic thinking tokens to validate parameter types, inspect returned error payloads, and self-correct across multi-turn agent loops with 99.7% execution reliability.

Related AI Architecture & Systems Guides

• - Computer-Use & Browser Agents: Automating Web Portals with AI Vision
• - How to Use Hybrid Reasoning in Gemini 3.7 Flash: Dynamic Thinking Budgets
• - AI Agent Tool Use & Safety Guards: Preventing Hallucinated API Calls in 2026
• - DeepSeek-R1 vs. Gemini 3.7 Flash: Open-Weight Reasoning vs. API Hybrid Thinking

Written by Shadab Alam

Founder & Engineer

I build custom web systems, automated backend workflows, and scalable e-commerce infrastructure for growing businesses. Founder at CodXpert & Anterpreneur.

LinkedIn -> Twitter / X -> Work with me ->