Bạn build chatbot cho enterprise. Cần tools:
- Giải thích MCP giải quyết vấn đề gì (tool integration maintenance burden)
- Phân biệt MCP vs tool use thông thường
- Hiểu architecture: Client, Server, Tools/Prompts/Resources
- Biết tại sao MCP đang thành industry standard
Mở đầu: Scale khi có 50 integrations
- GitHub (list repos, PRs, issues)
- Slack (send msg, read channels)
- Jira (create ticket, update status)
- Google Drive (search, read docs)
- Gmail (send, search emails)
- Salesforce (query leads)
- ...
Mỗi integration = 10-50 tool schemas + functions. 50 integrations × 30 tools = 1500 tools bạn maintain.
Mỗi thay đổi API upstream (Slack update v2) → bạn update 30 tool functions.
Unsustainable.
MCP = "Language for tool providers"
Idea: Service providers (GitHub, Slack, etc.) implement MCP servers exposing their tools qua chuẩn chung. Bạn chỉ cần connect MCP client tới servers.
Before MCP:
┌──────────────────────┐
│ Your chatbot app │
│ │
│ - GitHub tool 1 │
│ - GitHub tool 2 │
│ - ...50 tools │ ← BẠN MAINTAIN
│ - Slack tool 1 │
│ - ...30 tools │ ← BẠN MAINTAIN
│ - ... │
└──────────────────────┘
After MCP:
┌──────────────────────┐ ┌──────────────────┐
│ Your chatbot app │ ──MCP──▶ │ GitHub MCP │
│ (just client) │ │ (official) │
│ │ ──MCP──▶ │ Slack MCP │
│ │ │ (official) │
│ │ ──MCP──▶ │ Jira MCP │
└──────────────────────┘ └──────────────────┘Anthropic, AWS, GitHub, Google, Slack, Notion đều release MCP servers chính thức. Bạn reuse, không build.
MCP architecture
┌────────────────────────────────────────────────┐
│ │
│ Your App (MCP Client) │
│ ┌──────────────────┐ │
│ │ MCP Client SDK │ │
│ └────────┬─────────┘ │
│ │ │
│ │ Protocol messages │
│ │ (stdio, HTTP, WebSocket) │
│ ▼ │
│ ┌──────────────────┐ │
│ │ MCP Server │ │
│ │ ┌────────────┐ │ │
│ │ │ Tools │ │ tools (functions) │
│ │ │ Prompts │ │ prompts (templates) │
│ │ │ Resources │ │ resources (data) │
│ │ └────────────┘ │ │
│ └──────────────────┘ │
│ │ │
│ ▼ │
│ External service │
│ (GitHub API, DB, files, ...) │
│ │
└────────────────────────────────────────────────┘3 core concepts
1. Tools
Function Claude can call. Tương tự tool use Module 5, nhưng served qua MCP.
2. Prompts
Pre-built, high-quality prompt templates server provide. User pick prompt thay vì viết from scratch.
3. Resources
Data sources MCP server expose. Files, database records, documents. Client có thể read.
Module 8 cover từng piece.
MCP vs plain tool use
| Plain tool use | MCP | |
|---|---|---|
| Schema | Bạn viết | Server provides |
| Implementation | Bạn viết | Server implements |
| Updates | Bạn maintain | Server provider updates |
| Reuse | Per-app | Any MCP client can use |
| Overhead | Low (1 app) | Medium (need MCP client) |
| Scale | Hard | Easy (add servers) |
- When MCP worth
- > 5 integrations, long-term maintenance, team scale.
- When plain OK
- 1-3 tools cụ thể, MVP, tightly coupled to app.
Common questions
Ai viết MCP servers?
- Service providers: GitHub, AWS, Google, Slack đã public MCP servers
- Community: OSS devs wrap APIs
- Bạn: custom business logic internal services
MCP khác gì API?
- API: endpoints generic (REST, GraphQL)
- MCP server: wraps API với AI-friendly tool schemas, structured for LLM consumption
MCP khác gì tool use?
- Tool use: cơ chế Claude gọi function
- MCP: where tool schemas + functions live
MCP tools → Claude vẫn dùng tool use pattern.
Ai invent MCP?
Anthropic. Công bố 11/2024. Open standard, cộng đồng contribute.
Ecosystem growing
(2026-04) đã có:
- 100+ official servers: GitHub, Slack, Notion, Gmail, PostgreSQL, Stripe, ...
- 500+ community servers: niche tools
- Multiple clients: Claude Desktop, Claude Code, custom apps
- SDKs: Python, TypeScript, Go, Rust
Registry: modelcontextprotocol.io
Use cases
Developer workflows
GitHub MCP + Filesystem MCP + Terminal MCP → Claude Code super power.
Enterprise assistant
Gmail + Slack + Jira + Google Drive MCP → single chatbot interface mọi tools.
Data analysis
Snowflake + Postgres + Google Sheets MCP → query + analyze across databases.
Custom business
Build internal MCP server cho CRM, ERP. Team dùng qua Claude Desktop hoặc custom chatbot.
Anti-patterns preview
Build MCP server cho 1 internal tool
Overhead not worth. Plain tool use OK.
Fix: MCP khi > 3 integrations hoặc tool sẽ shared cross-app.
Wrap existing API 1-to-1
MCP server thêm overhead. Nếu không AI-optimize, not adding value.
Fix: Design MCP tools theo AI use case, không phải mirror API.
Ignore security
MCP server = executable code. Untrusted server = security risk.
Fix: Only connect to trusted servers.
Áp dụng ngay
Bài tập 1: Explore registry (10 phút)
Visit modelcontextprotocol.io. Browse servers. Identify 3 potentially useful cho app của bạn.
Bài tập 2: Plan MCP architecture (10 phút)
Cho app bạn xây trong khóa:
- Tools cần: ___
- MCP server hiện có match không?
- Build custom hay reuse?
Tóm tắt bài học
- MCP = standard protocol cho AI-tool integration. Solve maintenance burden của 50+ integrations.
- Client-server architecture. Client = your app. Server = tool provider.
- 3 concepts: Tools (functions), Prompts (templates), Resources (data).
- MCP vs tool use: MCP is where tools live. Tool use is how Claude calls them.
- Growing ecosystem 2026: 100+ official servers, industry adopting as standard.