Multi-Agent A2A Project (Google ADK + MCP Server)

This project implements a multi-agent travel assistant system where each agent shares access to a centralized MCP server to perform web searches via the Tavily tool. Instead of giving each agent its own search engine, this design uses a single MCP server to handle search queries, making the system more efficient and scalable. It demonstrates how centralized infrastructure (MCP) can support decentralized agents in a LangChain-based environment.

Getting Started

1. Clone the Repository

git clone https://github.com/Vinh-Sta/Multi_agent_with_MCP_Sever.git
cd Multi_agent_with_MCP_Sever

2. Set Up the Environment

python3 -m venv adk_demo  
source adk_demo/bin/activate  
pip install -r requirements.txt

3. Configure API Keys

GOOGLE_API_KEY="your-key"  
AGENT_MODEL="your-model"  
TAVILY_API_KEY="YOUR_API_KEY"

4. Run the Agent and UI

Start each agent and the UI:

uvicorn agents.host_agent.__main__:app --port 8080 &  
uvicorn agents.flight_agent.__main__:app --port 8081 &  
uvicorn agents.stay_agent.__main__:app --port 8082 &  
uvicorn agents.activities_agent.__main__:app --port 8083 &  


streamlit run travel_ui.py

4. System Architecture Diagram

┌─────────────────────┐
│    Travel UI (user) │
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│    Host Agent       │◄────────────────────────────────┐
└────────┬────────────┘                                 │
         ▼                                              │
┌────────────┐   ┌────────────┐   ┌────────────┐   ┌────▼─────┐
│Flight Agent│   │ Stay Agent│   │Activity Agent│  │ MCP Server│
└────┬───────┘   └────┬──────┘   └────┬────────┘   └────┬─────┘
     │               │               │                 │
     └───────────────┴───────────────┴─────────────────┘
                      ▼
              🌐 Tavily Web Search

Each agent makes search requests via the shared MCP server, which handles communication with the Tavily API and returns results to the respective agent.

5. Documentation

All detailed theory, design principles, and presentation materials related to the Model Context Protocol (MCP) used in this project are compiled in the file:

Model_Context_Protocol.pdf

Please refer to this document for an in-depth understanding of MCP architecture, agent coordination, and communication protocols.