MCP Server Architect (Modern AI Tools & Agentic Protocol)

English | Bahasa Indonesia

---

<a name="english"></a>

English

Orchestration & Integration

• Upstream AI Models & Protocol: ai-llm-integration-expert, frontier-ai-models-expert, mcp-sampling-critic
• Agent Clients & Swarms: multi-agent-orchestration, parallel-tool-orchestrator, gemini-agent-booster
• Domain Server Implementations: doku-mcp-server, xerj-search-expert, database-orm-expert
• Security & Authorization: zero-trust-secret-vault, authentication-identity-expert, prompt-injection-firewall
• Lifecycle & Governance: zero-to-prod-orchestrator, session-memory-manager, anti-slop

Description

Ultimate architectural guide for engineering high-performance, production-ready AI Tools/Bots via the Model Context Protocol (MCP v1.x). Enforces the use of FastMCP (Python) and @modelcontextprotocol/sdk (TypeScript). Mandates strict security guardrails, schema validation, stateful resource streaming, and support for both Standard Stdio and Streamable HTTP / Server-Sent Events (SSE) transports.

Trigger Conditions

• Building an MCP server to expose tools, resources, or prompt templates to AI agents.
• Integrating backend APIs, file systems, or databases as MCP agent tools.
• Implementing stateful, real-time MCP servers (resource subscriptions, log tailing, live metrics).
• Securing and auditing MCP servers exposing sensitive financial or production data.

---

SDK Selection (Mandatory Standard)

1. Python: FastMCP (FastAPI-like high-level DX for MCP tools, resources, and prompt templates).
2. TypeScript: @modelcontextprotocol/sdk (official SDK using the McpServer high-level abstraction with zod).

---

Production Implementation Recipes

Recipe 1: Production TypeScript MCP Server (Streamable HTTP / SSE)

import { McpServer, ResourceTemplate } from '@modelcontextprotocol/sdk/server/mcp.js';
import { SSEServerTransport } from '@modelcontextprotocol/sdk/server/sse.js';
import express from 'express';
import { z } from 'zod';

// Initialize the high-level MCP Server
const server = new McpServer({
  name: 'enterprise-analytics-mcp',
  version: '1.0.0',
});

// Register a type-safe Tool with Zod validation
server.tool(
  'calculate_metrics',
  'Calculates business analytics metrics across timeframes',
  {
    metricName: z.enum(['arr', 'churn', 'cac', 'ltv']).describe('The metric to compute'),
    quarter: z.string().regex(/^Q[1-4]-202[0-9]$/).describe('Target quarter, e.g., Q1-2026'),
  },
  async ({ metricName, quarter }) => {
    // Implement business logic with database access
    const mockData = { arr: '$2.4M', churn: '1.2%', cac: '$450', ltv: '$9,200' };
    return {
      content: [
        {
          type: 'text',
          text: JSON.stringify({ metric: metricName, quarter, value: mockData[metricName] }),
        },
      ],
    };
  }
);

// Register a Resource Template with dynamic URI
server.resource(
  'system_health',
  new ResourceTemplate('system://health/{service}', { list: undefined }),
  async (uri, { service }) => {
    return {
      contents: [
        {
          uri: uri.href,
          text: JSON.stringify({ service, status: 'HEALTHY', latencyMs: 14, timestamp: new Date().toISOString() }),
        },
      ],
    };
  }
);

// Expose via Express with SSE Transport
const app = express();
let transport: SSEServerTransport | null = null;

app.get('/sse', async (req, res) => {
  transport = new SSEServerTransport('/messages', res);
  await server.connect(transport);
});

app.post('/messages', async (req, res) => {
  if (transport) {
    await transport.handlePostMessage(req, res);
  } else {
    res.status(400).send('Transport not established');
  }
});

app.listen(3001, () => {
  console.log('MCP Server listening on http://localhost:3001/sse');
});

Recipe 2: Production FastMCP Server (Python)

from fastmcp import FastMCP, Context
from pydantic import BaseModel, Field
from typing import Literal

mcp = FastMCP("enterprise-vault-mcp", dependencies=["pydantic"])

class QueryParams(BaseModel):
    account_id: str = Field(..., description="UUID of customer account")
    status_filter: Literal["active", "suspended", "all"] = Field("active", description="Status filter")

@mcp.tool(name="fetch_account_summary", description="Retrieves account telemetry and balance")
async def fetch_account_summary(params: QueryParams, ctx: Context) -> str:
    ctx.info(f"Auditing request for account: {params.account_id}")
    
    # Secure business logic with RLS validation
    result = {
        "account_id": params.account_id,
        "balance_usd": 125430.50,
        "tier": "enterprise",
        "status": params.status_filter
    }
    return str(result)

@mcp.resource("config://app-settings")
def get_app_settings() -> str:
    """Provides application configuration context to the agent."""
    return '{"environment": "production", "rate_limit_rpm": 600, "region": "ap-southeast-1"}'

if __name__ == "__main__":
    # Runs standard Stdio transport or streamable HTTP
    mcp.run(transport="stdio")

---

Security & Operational Guardrails

1. OAuth 2.1 & Bearer Authentication: Bind session tokens to transport connections. Validate claims before executing any tool logic.
2. Schema Strictness: Never use untyped payloads. Every argument must have explicit types, range constraints, and descriptions to guide LLM tool-calling accuracy.
3. Row-Level Security (RLS): Enforce tenant and user context propagation to the database layer.
4. Circuit Breakers & Rate Limits: Cap consecutive tool executions per agent turn to prevent endless agentic recursive loops.
5. Idempotency: All destructive or state-mutating tools must require an idempotency_key argument.

Agent MCP Client Consumption & Tool Discovery

When acting as an AI Agent consuming external MCP servers:

1. Dynamic Tool Discovery: Check list_resources or mcp_config.json before assuming external capabilities do not exist.
2. Defensive Schema Querying: Never guess database schema or table names. Always execute list_tables or get_schema before generating SQL queries (execute_sql).
3. Cross-System Workflow Loop: Dynamically chain tools across domains: GitHub MCP (find issue) -> grep_search (locate file) -> autonomous-tdd-debugger (test & fix) -> GitHub MCP (create PR).
4. Rate Limit Awareness: Avoid rapid unthrottled loops against external MCP servers.

High-Throughput / Batched MCP Tool Execution (Gemini 4 Pro Readiness & Fallback)

• Ensure your MCP endpoints can handle Massive Parallel Tool Execution as models shift towards Agentic MoE architectures (Gemini 4 Pro / Gemini 4 Argon, Claude Opus 5.5).
• Availability Fallback Invariant: When orchestrating multi-tool agent loops:
  • Automatically and seamlessly degrade tasks from Gemini 4 (Argon & Pro) to Gemini 3.8 Flash if Gemini 4 is unprovisioned, rate-limited, or unavailable (preserving native parallel tool call execution, sub-100ms TTFT, and 1M+ context window).
  • Automatically and seamlessly degrade tasks from Claude Opus 5.5 to Claude Opus 4.6 if unprovisioned or rate-limited.
  • Zero disruption: MCP tool dispatchers must execute immediately through fallback paths without throwing unhandled missing-model errors.
• Implement parallel sub-processing in tools that expect to be called concurrently (e.g., using Promise.all or asyncio.gather for batch database queries instead of blocking sequentially).

---

<a name="bahasa-indonesia"></a>

Bahasa Indonesia

Integrasi Orkestrasi

• Model AI & Protokol Hulu: ai-llm-integration-expert, frontier-ai-models-expert, mcp-sampling-critic
• Klien Agen & Swarm: multi-agent-orchestration, parallel-tool-orchestrator, gemini-agent-booster
• Implementasi Server Domain: doku-mcp-server, xerj-search-expert, database-orm-expert
• Keamanan & Otorisasi: zero-trust-secret-vault, authentication-identity-expert, prompt-injection-firewall
• Siklus Hidup & Tata Kelola: zero-to-prod-orchestrator, session-memory-manager, anti-slop

Deskripsi

Panduan arsitektur utama untuk membangun AI Tools/Bots modern dan siap produksi via Model Context Protocol (MCP v1.x). Mewajibkan penggunaan FastMCP (Python) dan @modelcontextprotocol/sdk (TypeScript). Menerapkan pengamanan ketat, validasi skema, streaming resource stateful, serta dukungan transport Standar Stdio maupun Streamable HTTP / SSE.

Kondisi Pemicu

• Membangun server MCP untuk mengekspos alat (tools), resource, atau template prompt ke agen AI.
• Mengintegrasikan API backend, sistem file, atau database sebagai alat agen AI.
• Mengimplementasikan server MCP stateful dan real-time (langganan resource, tailing log, metrik langsung).
• Mengamankan server MCP yang mengekspos data finansial atau produksi yang sensitif.

---

Standar Pemilihan SDK (Wajib)

1. Python: FastMCP (pengalaman developer tingkat tinggi ala FastAPI untuk tools, resource, dan template prompt).
2. TypeScript: @modelcontextprotocol/sdk (SDK resmi menggunakan abstraksi McpServer dengan zod).

---

Resep Implementasi Produksi

Resep 1: Server MCP TypeScript Produksi (Streamable HTTP / SSE)

import { McpServer, ResourceTemplate } from '@modelcontextprotocol/sdk/server/mcp.js';
import { SSEServerTransport } from '@modelcontextprotocol/sdk/server/sse.js';
import express from 'express';
import { z } from 'zod';

const server = new McpServer({
  name: 'enterprise-analytics-mcp',
  version: '1.0.0',
});

// Pendaftaran Tool dengan validasi ketat Zod
server.tool(
  'calculate_metrics',
  'Menghitung metrik analitik bisnis untuk kuartal tertentu',
  {
    metricName: z.enum(['arr', 'churn', 'cac', 'ltv']).describe('Metrik yang ingin dihitung'),
    quarter: z.string().regex(/^Q[1-4]-202[0-9]$/).describe('Target kuartal, misal: Q1-2026'),
  },
  async ({ metricName, quarter }) => {
    const data = { arr: '$2.4M', churn: '1.2%', cac: '$450', ltv: '$9,200' };
    return {
      content: [
        {
          type: 'text',
          text: JSON.stringify({ metrik: metricName, kuartal: quarter, nilai: data[metricName] }),
        },
      ],
    };
  }
);

// Pendaftaran Template Resource dengan URI Dinamis
server.resource(
  'system_health',
  new ResourceTemplate('system://health/{service}', { list: undefined }),
  async (uri, { service }) => {
    return {
      contents: [
        {
          uri: uri.href,
          text: JSON.stringify({ layanan: service, status: 'HEALTHY', latensiMs: 14, waktu: new Date().toISOString() }),
        },
      ],
    };
  }
);

const app = express();
let transport: SSEServerTransport | null = null;

app.get('/sse', async (req, res) => {
  transport = new SSEServerTransport('/messages', res);
  await server.connect(transport);
});

app.post('/messages', async (req, res) => {
  if (transport) {
    await transport.handlePostMessage(req, res);
  } else {
    res.status(400).send('Transport belum terhubung');
  }
});

app.listen(3001, () => {
  console.log('Server MCP berjalan pada http://localhost:3001/sse');
});

Resep 2: Server FastMCP Produksi (Python)

from fastmcp import FastMCP, Context
from pydantic import BaseModel, Field
from typing import Literal

mcp = FastMCP("enterprise-vault-mcp", dependencies=["pydantic"])

class ParameterAkun(BaseModel):
    account_id: str = Field(..., description="UUID akun pengguna")
    status_filter: Literal["active", "suspended", "all"] = Field("active", description="Filter status")

@mcp.tool(name="ambil_ringkasan_akun", description="Mengambil telemetri dan saldo akun")
async def ambil_ringkasan_akun(params: ParameterAkun, ctx: Context) -> str:
    ctx.info(f"Memproses permintaan untuk akun: {params.account_id}")
    hasil = {
        "account_id": params.account_id,
        "saldo_usd": 125430.50,
        "tier": "enterprise",
        "status": params.status_filter
    }
    return str(hasil)

@mcp.resource("config://app-settings")
def ambil_pengaturan_aplikasi() -> str:
    """Menyediakan konteks konfigurasi aplikasi ke agen AI."""
    return '{"environment": "production", "rate_limit_rpm": 600, "region": "ap-southeast-1"}'

if __name__ == "__main__":
    mcp.run(transport="stdio")

---

Keamanan & Batasan Operasional

1. Otentikasi OAuth 2.1 & Bearer: Ikat token sesi ke koneksi transport. Validasi hak akses sebelum mengeksekusi logika alat.
2. Validasi Skema Ketat: Hindari penggunaan parameter tanpa tipe data yang jelas. Setiap argumen wajib memiliki tipe data, batas nilai, dan deskripsi.
3. Row-Level Security (RLS): Teruskan identitas pengguna dan penyewa (tenant) ke lapisan database driver.
4. Circuit Breakers & Rate Limits: Batasi pemanggilan tool berulang dalam satu giliran respon untuk mencegah perulangan tak terkontrol (infinite loops).
5. Idempotency: Semua tool yang memodifikasi data wajib mendukung argumen idempotency_key.

Konsumsi Klien MCP & Eksplorasi Tool oleh Agen

Ketika agen bertindak sebagai Klien MCP:

1. Eksplorasi Tool Dinamis: Periksa list_resources atau konfigurasi MCP sebelum menyimpulkan kapabilitas tidak tersedia.
2. Kueri Skema Defensif: Jangan pernah menebak nama tabel/skema. Selalu gunakan list_tables atau get_schema sebelum membuat kueri SQL.
3. Alur Kerja Lintas Sistem: Rangkaikan pemanggilan tool antar-domain: GitHub MCP -> pencarian kode lokal -> perbaikan otonom -> Pull Request GitHub.
4. Kesadaran Batas Frekuensi: Hindari loop pemanggilan berulang tanpa jeda waktu saat memanggil server MCP eksternal.

Eksekusi Tool MCP Batch / Throughput Tinggi (Kesiapan Gemini 4 Pro & Fallback)

• Pastikan endpoint MCP Anda mampu menangani Eksekusi Tool Paralel Masif seiring transisi model menuju arsitektur Agentic MoE (Gemini 4 Pro / Gemini 4 Argon, Claude Opus 5.5).
• Invarian Pengalihan Ketersediaan (Availability Fallback Invariant): Ketika mengorkestrasi loop agen multi-tool:
  • Alihkan tugas dari Gemini 4 (Argon & Pro) secara otomatis ke Gemini 3.8 Flash jika Gemini 4 belum provisioned, terkena batas kuota, atau tidak tersedia (mempertahankan eksekusi tool paralel native, latensi sub-100ms TTFT, dan jendela konteks 1M+).
  • Alihkan tugas dari Claude Opus 5.5 secara otomatis ke Claude Opus 4.6 jika belum provisioned atau terbatas kuota.
  • Tanpa interupsi: Dispatcher tool MCP harus segera dieksekusi melalui jalur fallback tanpa galat model tidak ditemukan yang tidak tertangani.
• Terapkan pemrosesan sub-tugas paralel pada tools yang mungkin dipanggil secara bersamaan (misal: gunakan Promise.all atau asyncio.gather untuk query batch, alih-alih mengeblok secara berurutan).