MCP代理系统

┌─────────────────┐
│  User Request   │
└────────┬────────┘
         │
         ▼
┌─────────────────────┐
│ Coordinator Agent   │  ◄── Plans & Decides
│   (Ollama LLM)      │
└─────────┬───────────┘
          │ A2A Communication
          ▼
┌─────────────────────┐
│   Worker Agent      │  ◄── Executes Tasks
│   (Ollama LLM)      │
└─────────┬───────────┘
          │
          ▼
┌─────────────────────┐
│    MCP Server       │  ◄── Tool Execution
│  (weather, time,    │
│   calculator)       │
└─────────────────────┘

mkdir mcp-agents-demo
cd mcp-agents-demo

python3 -m venv venv
source venv/bin/activate  # On Windows: venv/Scripts/activate

pip install mcp httpx

# Install Ollama (if not already installed)
# curl -fsSL https://ollama.com/install.sh | sh

# Start Ollama server
ollama serve

# In another terminal, pull a model
# ollama pull llama2  # or llama3, mistral, etc.

python3 agents.py

mcp-agents-demo/
├── mcp_server.py       # MCP server with tools
├── agents.py           # Two-agent system
├── requirements.txt    # Python dependencies
└── README.md          # This file

Responsibilities:
- Analyze user requests
- Understand available tools
- Create execution plans
- Delegate to Worker Agent

User: "What's the weather in Tokyo?"
↓
Coordinator: Analyzes → Identifies "get_weather" tool needed
↓
Delegates to Worker with plan

Responsibilities:
- Receive tasks from Coordinator
- Select appropriate MCP tools
- Execute tool calls
- Return results

Receives: "Get weather for Tokyo"
↓
Calls: get_weather(city="Tokyo")
↓
Returns: Weather data

# Example A2A flow
coordinator_result = await coordinator.process(request)
↓
worker_result = await worker.execute_task(
    coordinator_result["task"],
    coordinator_result["decision"]  # A2A context sharing
)

OLLAMA_MODEL = "llama3"  # Options: llama2, llama3, mistral, codellama

test_requests = [
    "What is the weather in Tokyo?",
    "Calculate 15 divided by 3",
    "What time is it in Paris?",
    "Your custom request here"
]

types.Tool(
    name="your_tool",
    description="Tool description",
    inputSchema={
        "type": "object",
        "properties": {
            "param": {"type": "string"}
        },
        "required": ["param"]
    }
)

elif name == "your_tool":
    result = your_implementation(arguments)
    return [types.TextContent(type="text", text=json.dumps(result))]

======================================================================
🚀 Starting MCP Agent System (A2A Demo)
======================================================================

✅ Connected to MCP server

======================================================================
📝 TEST CASE 1: What is the weather in Tokyo?
======================================================================

🎯 COORDINATOR AGENT: Analyzing request...
   Request: What is the weather in Tokyo?

🤖 Calling Ollama...

📋 Coordinator Analysis:
   This request requires the get_weather tool...

----------------------------------------------------------------------

🔄 AGENT-TO-AGENT COMMUNICATION
   Coordinator → Worker: Delegating task

⚙️  WORKER AGENT: Executing task...
   Task: What is the weather in Tokyo?

📡 Calling MCP tool: get_weather(city='Tokyo')
✅ Tool Result: {
  "city": "Tokyo",
  "temperature": 23,
  "condition": "Sunny",
  "humidity": 65
}

======================================================================
✅ FINAL RESULT:
{
  "success": true,
  "result": "...",
  "tool": "get_weather"
}
======================================================================

python3 agents.py

async def interactive_mode():
    mcp_client = MCPClientWrapper()
    await mcp_client.connect()
    
    coordinator = CoordinatorAgent(mcp_client)
    worker = WorkerAgent(mcp_client)
    
    print("Interactive mode. Type 'quit' to exit.\n")
    
    while True:
        user_input = input("Your request: ")
        if user_input.lower() in ['quit', 'exit']:
            break
        
        coord_result = await coordinator.process(user_input)
        worker_result = await worker.execute_task(
            coord_result["task"],
            coord_result["decision"]
        )
        
        print("\n✅ Result:")
        print(json.dumps(worker_result, indent=2))
        print()
    
    await mcp_client.close()

if __name__ == "__main__":
    asyncio.run(interactive_mode())

# Start Ollama in a separate terminal
ollama serve

# Verify it's running
curl http://localhost:11434/api/tags

# Activate virtual environment
source venv/bin/activate

# Install/reinstall dependencies
pip install --upgrade mcp httpx

# List available models
ollama list

# Pull the model you need
ollama pull llama2
# or
ollama pull llama3

export MCP_DEBUG=1
python3 agents.py

# Edit agents.py to change OLLAMA_MODEL
# Then run specific tests
python3 agents.py

import time

start = time.time()
result = await coordinator.process(request)
end = time.time()

print(f"Coordinator took {end - start:.2f}s")