Chapter 6 — MCP Ecosystem

🔌

"Anthropic opened MCP in Nov 2024.In less than 18 months: 97M downloads/month, 78% enterprise adoption.This is how AI 'plugs into' enterprise systems."

You'll learn

What MCP is + why Anthropic "won the standard"
78% enterprise has ≥1 MCP agent in production
A2A protocol (Google donated Linux Foundation) — parallel protocol
Blue ocean: build MCP for local stacks (Vietnamese, Brazilian, Indonesian)
How to build MCP server quickly

01 What is MCP?

Model Context Protocol = open standard for LLM agents to connect with external tools / data sources.

Pre-MCP problem

Each LLM vendor had own way:

OpenAI: function calling
Anthropic: tool use
Google: function declarations
Cursor, Windsurf, Claude Code: own integrations

→ N×M problem: N LLMs × M tools = chaos integration.

MCP solves

Before MCP:
LLM A → custom adapter → Tool 1
LLM A → custom adapter → Tool 2
LLM B → custom adapter → Tool 1
LLM B → custom adapter → Tool 2
(N × M)

After MCP:
LLM A ┐
LLM B ├ → MCP standard → Tool 1, Tool 2, Tool 3
LLM C ┘
(N + M)

Components

LLM (client) ↔ MCP standard ↔ MCP server
                                  - Tools
                                  - Resources
                                  - Prompts
                                  │
                                  ▼
                            [Service: GitHub / Slack / DB / etc.]

02 Insane stats (May 2026)

Metric	Number
MCP SDK downloads/month	97 million (Mar 2026)
Growth in 18 months	970x (from 100K in month 1)
MCP servers registered	17,468 (cross-registry census)
Official registry	5,800+
78% enterprise has ≥1 MCP agent in production	WorkOS report
67% CTOs name MCP default agent-integration	(vs A2A 23%, ACP 8%)

03 Adoption — who has adopted MCP?

LLM vendors

Vendor	Status
Anthropic	Creator, native
OpenAI	Apr 2025 (ChatGPT Apps SDK)
Google	Mar 2026 (Gemini API + Vertex AI Agent Builder)
Microsoft / Copilot	Partial (competing with own protocol)

IDE / coding tools

Tool	MCP support
Cursor
Windsurf
Zed
JetBrains
Claude Code	Native
Vercel AI SDK

Frameworks

Framework	MCP integration
LangChain / LangGraph
CrewAI
OpenAI Agents SDK

04 A2A protocol — competitor / complement

Google A2A (Agent-to-Agent)

Item	Detail
Announce	Apr 9, 2025
Donated to	Linux Foundation Jun 2025
Supporters	150+ — Atlassian, Salesforce, ServiceNow, SAP, Workday
Protocol	HTTP + SSE + JSON-RPC 2.0 + Agent Cards
Use case	Agent ↔ agent comm (different vendors)

MCP vs A2A — not conflicting

	MCP	A2A
Purpose	LLM ↔ tool/data	Agent ↔ agent
Standard owner	Anthropic	Google → Linux Foundation
Mature stage (May 2026)	Established (78% enterprise)	Early adoption
Best for	Single agent + many tools	Multi-vendor agent network

→ Learn MCP first, A2A later (when cross-vendor agents needed).

05 MCP server ecosystem

Top MCP servers (May 2026)

Category	Server	Use
Dev tools	github, postgres, sqlite, filesystem, git	Code + data ops
Cloud	aws, gcp, cloudflare, vercel	Infra automation
Productivity	slack, notion, linear, jira, asana	Work management
Customer / CRM	hubspot, salesforce, intercom	Sales/CS ops
Communication	gmail, outlook, calendar	Schedule + email
Analytics	google-analytics, amplitude, mixpanel	Data analysis
Design	figma, canva	Design ops
Browser	playwright, puppeteer	Web automation

Gateway / aggregator

Smithery — central registry + browser
Obot — enterprise MCP gateway
Webrix — multi-tenant MCP proxy
mcp.so — community discovery

06 Build MCP server — quickstart

Setup 5 minutes

Option 1: Use existing SDK

bash

# Python
pip install mcp

# TypeScript
npm install @modelcontextprotocol/sdk

Minimal server (TypeScript)

typescript

import { Server } from '@modelcontextprotocol/sdk/server/index.js';
import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';

const server = new Server({
  name: 'my-local-tools',
  version: '1.0.0',
}, {
  capabilities: { tools: {} },
});

server.setRequestHandler('tools/list', async () => ({
  tools: [{
    name: 'check_invoice',
    description: 'Check invoice status in accounting system',
    inputSchema: {
      type: 'object',
      properties: {
        invoice_id: { type: 'string' },
      },
      required: ['invoice_id'],
    },
  }],
}));

server.setRequestHandler('tools/call', async (req) => {
  if (req.params.name === 'check_invoice') {
    const result = await callAccountingAPI(req.params.arguments.invoice_id);
    return { content: [{ type: 'text', text: JSON.stringify(result) }] };
  }
});

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

Use in Claude Code / Cursor

json

// ~/.claude.json
{
  "mcpServers": {
    "local-tools": {
      "command": "node",
      "args": ["/path/to/my-local-tools/dist/server.js"],
      "env": { "API_KEY": "..." }
    }
  }
}

Restart Claude Code → tool mcp__local-tools__check_invoice appears

07 Blue ocean — MCP for local stacks

Current state by region

Region	Local stack	MCP server existing?
Vietnam	MISA, KiotViet, Sapo, Pancake, Base.vn, Misa AMIS	None yet
Indonesia	Mekari, Sleekr, OY!, Mokapos	None yet
India	Tally, Razorpay, Tata, Kotak	Minimal
Brazil	Pagar.me, NuBank, Conta Azul	Minimal
Philippines	GCash, Maya, Sprout HR	None yet

→ 100% blue ocean. Opportunity to build "MCP for [region] stack" as early winner.

Project ideas

Idea 1: `mcp-misa` (Vietnam)

Tool: check invoice, create voucher, generate VAT report
Target: agencies doing CS / accounting for VN SMEs
Revenue: open-source + service consulting

Idea 2: `mcp-kiotviet` (Vietnam)

Tool: check stock, create order, customer lookup
Target: F&B / retail using KiotViet
Revenue: license $50-200/month per business

Idea 3: `mcp-pancake` (Vietnam)

Tool: list conversations, send messages, update customer tag
Target: agencies using Pancake for clients
Revenue: subscription

Idea 4: `mcp-mekari` (Indonesia)

Tool: HRIS + payroll automation
Target: agencies for Indonesian SMEs
Revenue: open-source build trust

Idea 5: `mcp-local-payment` (region-specific)

VNPay + MoMo + ZaloPay (Vietnam)
Pix + PicPay (Brazil)
GCash + Maya (Philippines)

Go-to-market

Step	Action
1	Build MCP open-source (GitHub)
2	Submit to Smithery + mcp.so
3	Twitter / Reddit / LinkedIn launch post
4	Reach out to local dev communities
5	Speak at AI meetups
6	Build paid tier (hosted, support)

→ Become "MCP-for-[region]" go-to person — establish authority + lead inflow.

08 Enterprise use cases — MCP-driven workflows

Enterprise patterns

CS multi-system

Stack: Claude + mcp-pancake + mcp-misa + mcp-shopify

Customer message via Pancake
Agent: check order in Shopify, check invoice MISA
Reply with full context
Update CRM Pancake

Sales lead enrichment

Stack: Claude + mcp-hubspot + mcp-builtwith + mcp-google

Lead enters HubSpot
Agent: enrich from BuiltWith (tech stack) + Google (company info)
Score lead, route to sales rep

Inventory rebalance

Stack: Claude + mcp-kiotviet + mcp-sapo (multi-store)

Daily check stock across stores
Agent: suggest transfers between stores
Auto-create transfer order

HR onboarding

Stack: Claude + mcp-base.vn + mcp-slack + mcp-google

New employee → create account in Base.vn + Slack + Google Workspace
Send welcome email + checklist

09 Common pitfalls

🚨 6 MCP server dev mistakes

1. Unclear tool names → agent doesn't pick. Use namespace service_action (e.g., misa_invoice_check)

2. Loose schema → agent generates wrong inputs. Validate strict with Zod / Pydantic

3. Token-inefficient output → bloats context. Paginate, truncate, filter

4. Unclear error messages → agent retries infinitely. Return error code + suggest fix

5. Skip auth / security → MCP server leaks data. Per-user auth + audit log

6. No eval testing → tool works happy path, fails edge. Test 20+ scenarios

10 Roadmap for dev to MCP expert

6 months → MCP expert + service business

Month 1: Learn MCP basics

Build 3 hello-world servers (filesystem, HTTP, DB)
Read Anthropic docs + best practices

Month 2: First local MCP

Pick 1 local stack
Build full MCP server for 1 use case
Launch GitHub open-source

Month 3: Distribution

Submit registry (Smithery + mcp.so)
Blog post, Twitter thread, demo video
Speak at meetups

Month 4: Second + third MCP

Add complementary stack
Cross-promote with first MCP

Month 5: Service business

Pitch 3 SMEs: full MCP-driven automation
Charge $5-15K project

Month 6: Recurring + scale

Hosted tier ($50-200/month per server per business)
Speak at conferences, build authority

11 Practice exercises

✍️ 3 levels

Level 1 — 1 week

Setup MCP SDK (TS or Python)
Build hello-world: tool echo
Connect Claude Code, test

Level 2 — 1 month

Pick 1 local service
Build MCP server with 5 tools
Open-source GitHub + Smithery

Level 3 — 6 months

3 production MCP servers for local stacks
5 paying customers (subscription)
$1-3K MRR

12 Continue reading

Final word

"MCP is USB-C for AI agents.Before MCP: each vendor had own port.After MCP: 1 port plugs everywhere.Emerging markets have no MCP for local stacks yet.Whoever builds first = wins the category.The door is open. Walk through or watch others — your choice."

Chapter 6 — MCP Ecosystem ​

01 What is MCP? ​

Pre-MCP problem ​

MCP solves ​

Components ​

02 Insane stats (May 2026) ​

03 Adoption — who has adopted MCP? ​

LLM vendors ​

IDE / coding tools ​

Frameworks ​

04 A2A protocol — competitor / complement ​

Google A2A (Agent-to-Agent) ​

MCP vs A2A — not conflicting ​

05 MCP server ecosystem ​

Top MCP servers (May 2026) ​

Gateway / aggregator ​

06 Build MCP server — quickstart ​

Option 1: Use existing SDK ​

Minimal server (TypeScript) ​

Use in Claude Code / Cursor ​

07 Blue ocean — MCP for local stacks ​

Current state by region ​

Project ideas ​

Idea 1: mcp-misa (Vietnam) ​

Idea 2: mcp-kiotviet (Vietnam) ​

Idea 3: mcp-pancake (Vietnam) ​

Idea 4: mcp-mekari (Indonesia) ​

Idea 5: mcp-local-payment (region-specific) ​

Go-to-market ​

08 Enterprise use cases — MCP-driven workflows ​

CS multi-system ​

Sales lead enrichment ​

Inventory rebalance ​

HR onboarding ​

09 Common pitfalls ​

10 Roadmap for dev to MCP expert ​

11 Practice exercises ​

12 Continue reading ​