How to Build an AI Agent with Model Context Protocol (MCP) — Step by Step

How to Build an AI Agent with Model Context Protocol (MCP)

Sun Jul 19 2026 · tidqom.com engineering

MCP has quietly become the default way agents talk to tools in 2026.

What is MCP

Related: Midjourney vs DALL-E vs Flux: A Practical Image-Tool Comparison Framework →

Related: Anthropic Expands Claude Cowork to Mobile and Web, Broadening the Agent Surface →

A small JSON-RPC protocol that lets any LLM discover and call tools exposed by any server. Think USB-C for AI agents.

Step 1: Design tools

Related: 9 Free AI Coding Tools Every Developer Should Try in 2026 →

Related: Claude 4.5 vs GPT-5.5: Which AI Coding Model Wins in 2026? →

• Clear names (create_invoice, not do_stuff).
• Narrow JSON schemas.
• Compact, serializable results.
• Descriptions the model reads at planning time.

Step 2: Build the server

Related: Sora 2 Review: OpenAI's Video Model Is Finally Useful for Real Work →

Related: الذكاء الاصطناعي الوكيلي (Agentic AI): الدليل الشامل 2026 →

import { McpServer } from "@modelcontextprotocol/sdk/server";
import { z } from "zod";

const server = new McpServer({ name: "invoicing", version: "1.0.0" }); server.tool("create_invoice", { customerId: z.string(), amountCents: z.number().int().positive() }, async ({ customerId, amountCents }) => { const inv = await db.invoices.create({ customerId, amountCents }); return { content: [{ type: "text", text: JSON.stringify(inv) }] }; }); server.listen();

Step 3: Wire the model

Related: How to Clone Your Voice with AI in 2026 (Free and Paid Options) →

Related: Anthropic Extends Free Claude Fable 5 Access Through July 19 — Again →

Loop: model plans → calls tool → server executes → model incorporates → repeat until stepCountIs(50).

Step 4: Memory

Related: Anthropic Localizes Claude Pricing for India, Its Biggest Market After the US →

• Short-term buffer with summarization at 20k tokens.
• Long-term vector store retrieved via search_memory tool.

Step 5: Harden

Related: Anthropic's Claude Moves Into Healthcare Claims and Care Management →

• Auth on every tool.
• Rate limits per user.
• Hard cost caps.
• Approval gates for spend/send.
• Idempotency keys.

Step 6: Evaluate

30–100 real tasks. Score on correctness, efficiency, safety. Regression-test on every change.

Pitfalls

1. Descriptions too long.
2. Overlapping tools.
3. No approval gates.
4. Model-side authorization.
5. No cost caps.

FAQ

MCP vs function calling? MCP standardizes so any client can use any server. Which SDK? TypeScript for edge/serverless; Python for ML stacks. Open-source models? Llama 4, DeepSeek R2, Qwen 3 all support tool calling. Secure? Transport only — auth, limits, approvals, audits are your job.

Related

• Blog
• AI Models
• AI Tools

Related reading

• Anthropic prompt generator guide
• Devin AI review
• Are AI coding agents replacing developers?
• Explore the AI models directory and the blog.