MCP là gì?

Định nghĩa

MCP (Model Context Protocol) là một open-source standard cho phép kết nối AI applications với external systems. MCP cung cấp cách chuẩn hóa để AI agents truy cập data sources, tools, và workflows.

Analogy: USB-C cho AI

USB-C cho devices:
- Standardized connector
- Power delivery
- Data transfer
- Works across devices

MCP cho AI:
- Standardized protocol
- Tool execution
- Data access
- Works across AI platforms

Vấn đề MCP giải quyết

Before MCP

AI Agent trong "sandbox":
┌─────────────────────┐
│   AI Agent          │
│   - No file access  │
│   - No database     │
│   - No APIs         │
│   - Limited context │
└─────────────────────┘

Problems:

❌ Cannot read local files
❌ Cannot query databases
❌ Cannot call external APIs
❌ Cannot integrate with existing systems
❌ Each AI platform has different integration

After MCP

MCP enables:
┌─────────────┐    MCP    ┌─────────────────┐
│ AI Agent    │◄────────►│ External Systems│
│             │          │ • Databases     │
│ • Claude    │          │ • Files         │
│ • Cursor    │          │ • APIs          │
│ • ChatGPT   │          │ • Tools         │
└─────────────┘          └─────────────────┘

Benefits:

✅ Standardized access to external systems
✅ Security model built-in
✅ Works across AI platforms
✅ Rich ecosystem of servers
✅ Extensible and customizable

MCP trong AI Ecosystem

Position trong Stack

┌─────────────────────────┐
│    AI Applications      │
│  (Claude, Cursor, etc.) │
└────────────┬────────────┘
             │ MCP Protocol
┌────────────▼────────────┐
│     MCP Servers         │
│  • Database servers     │
│  • File servers         │
│  • API servers          │
│  • Custom servers       │
└────────────┬────────────┘
             │
┌────────────▼────────────┐
│    External Systems     │
│  • MongoDB              │
│  • PostgreSQL           │
│  • File systems         │
│  • REST APIs            │
└─────────────────────────┘

Comparison với các protocols khác

Protocol	Purpose	Scope	AI-specific
HTTP/REST	Web APIs	General	No
gRPC	Microservices	General	No
WebSocket	Real-time	General	No
LSP	Language services	Development	Partial
MCP	AI context	AI-specific	✅ Yes

Core Concepts

1. Client

AI application sử dụng MCP:

Claude Desktop
Cursor IDE
Custom AI applications
Enterprise AI platforms

2. Server

Cung cấp functionality qua MCP:

Database servers
File servers
API servers
Custom business logic servers

3. Transport

Cách client và server communicate:

stdio: Standard input/output (default)
HTTP: HTTP-based communication
Custom: Custom transport protocols

4. Protocol

JSON-RPC 2.0 based protocol:

Standardized message format
Capability negotiation
Error handling
Extensibility

MCP Components Deep Dive

Tools

Definition: Functions AI có thể actively gọi

Characteristics:

AI decides khi nào dùng
Can modify external systems
Return results cho AI
Can be chained together

Examples:

javascript

// Database query tool
{
  "name": "query_database",
  "description": "Execute SQL query",
  "inputSchema": {
    "type": "object",
    "properties": {
      "query": {"type": "string"},
      "database": {"type": "string"}
    }
  }
}

// File write tool
{
  "name": "write_file",
  "description": "Write content to file",
  "inputSchema": {
    "type": "object",
    "properties": {
      "path": {"type": "string"},
      "content": {"type": "string"}
    }
  }
}

Resources

Definition: Passive data sources cho context

Characteristics:

Read-only access
Provide context cho AI
Can be subscribed to for updates
AI cannot modify

Examples:

javascript

// File resource
{
  "uri": "file:///path/to/config.json",
  "name": "Configuration File",
  "description": "Application configuration",
  "mimeType": "application/json"
}

// Database schema resource
{
  "uri": "mongodb://localhost/users/schema",
  "name": "Users Collection Schema",
  "description": "Schema definition for users collection",
  "mimeType": "application/json"
}

Prompts

Definition: Pre-built instruction templates

Characteristics:

Guide AI behavior
Combine tools và resources
User-controlled
Reusable templates

Examples:

javascript

// Data analysis prompt
{
  "name": "analyze_database",
  "description": "Analyze database and provide insights",
  "arguments": [
    {
      "name": "database_schema",
      "description": "Database schema resource",
      "required": true
    },
    {
      "name": "analysis_goal",
      "description": "What to analyze",
      "required": true
    }
  ]
}

Use Cases

1. Database Integration

AI Agent: "Show me all users who signed up last week"
MCP Server: 
- Connect to database
- Execute query
- Return results

2. File System Access

AI Agent: "Read the configuration file and update the timeout"
MCP Server:
- Read config file (resource)
- Update file (tool)
- Confirm changes

3. API Integration

AI Agent: "Create a new user in Stripe"
MCP Server:
- Call Stripe API (tool)
- Handle authentication
- Return user ID

4. Development Workflow

AI Agent: "Run tests and fix any failures"
MCP Server:
- Run tests (tool)
- Read test results (resource)
- Fix code (tool)
- Rerun tests (tool)

MCP vs Custom Integrations

Custom Integration (trước MCP)

javascript

// Custom Claude plugin
class ClaudePlugin {
  async connectToDatabase() {
    // Custom implementation
    // Hard-coded for Claude
    // No standardization
  }
  
  async readFile() {
    // Custom implementation
    // Security concerns
    // Limited functionality
  }
}

// Custom Cursor plugin
class CursorPlugin {
  async connectToDatabase() {
    // Different implementation
    // Different API
    // Duplication of effort
  }
}

MCP Integration (sau MCP)

javascript

// Single MCP server
class DatabaseMCPServer {
  async listTools() {
    return [
      {
        name: "query",
        description: "Execute database query"
      }
    ];
  }
  
  async callTool(name, args) {
    if (name === "query") {
      return await this.executeQuery(args.query);
    }
  }
}

// Works with Claude, Cursor, ChatGPT, etc.
// Standardized protocol
- Single implementation
- Consistent behavior
- Security built-in

MCP Philosophy

1. Standardization

Cùng protocol cho tất cả AI platforms
Consistent behavior
Reduced development effort

2. Security First

Authentication và authorization
Permission model
Audit trail

3. Extensibility

Easy to add new servers
Custom business logic
Community contributions

4. Interoperability

Cross-platform compatibility
Mixed server environments
Gradual adoption

MCP Timeline

2024 Q1: Anthropic introduces MCP
2024 Q2: Open source protocol
2024 Q3: First servers released
2024 Q4: Major IDE integrations
2025 Q1: Enterprise adoption
2025 Q2: Ecosystem growth

Key Players

Maintainers

Anthropic (creator)
OpenAI (contributor)
Block (contributor)
Linux Foundation (host)

Major Implementations

Claude Desktop (native)
Cursor IDE (native)
MongoDB (official server)
Community servers (100+)

Adoption

50+ AI applications support MCP
200+ MCP servers available
Growing enterprise adoption

Summary

MCP là:

✅ Standard protocol cho AI integration
✅ Solves real problems với AI agents
✅ Growing ecosystem
✅ Production-ready

MCP không phải là:

❌ Magic solution cho tất cả problems
❌ Replacement cho existing protocols
❌ Only for databases
❌ Limited to specific AI platforms

Core value:

"MCP enables AI agents to safely and effectively interact with external systems through a standardized protocol"

Tiếp theo: 2. Kiến trúc MCP → Hiểu sâu về cách MCP hoạt động

MCP là gì? ​

Định nghĩa ​

Analogy: USB-C cho AI ​

Vấn đề MCP giải quyết ​

Before MCP ​

After MCP ​

MCP trong AI Ecosystem ​

Position trong Stack ​

Comparison với các protocols khác ​

Core Concepts ​

1. Client ​

2. Server ​

3. Transport ​

4. Protocol ​

MCP Components Deep Dive ​

Tools ​

Resources ​

Prompts ​

Use Cases ​

1. Database Integration ​

2. File System Access ​

3. API Integration ​

4. Development Workflow ​

MCP vs Custom Integrations ​

Custom Integration (trước MCP) ​

MCP Integration (sau MCP) ​

MCP Philosophy ​

1. Standardization ​

2. Security First ​

3. Extensibility ​

4. Interoperability ​

MCP Timeline ​

Key Players ​

Maintainers ​

Major Implementations ​

Adoption ​

Summary ​

MCP là gì?

Định nghĩa

Analogy: USB-C cho AI

Vấn đề MCP giải quyết

Before MCP

After MCP

MCP trong AI Ecosystem

Position trong Stack

Comparison với các protocols khác

Core Concepts

1. Client

2. Server

3. Transport

4. Protocol

MCP Components Deep Dive

Tools

Resources

Prompts

Use Cases

1. Database Integration

2. File System Access

3. API Integration

4. Development Workflow

MCP vs Custom Integrations

Custom Integration (trước MCP)

MCP Integration (sau MCP)

MCP Philosophy

1. Standardization

2. Security First

3. Extensibility

4. Interoperability

MCP Timeline

Key Players

Maintainers

Major Implementations

Adoption

Summary