Strands AI Agents with MCP and Neo4j

pip install strands mcp neo4j pandas matplotlib seaborn openai

pip install uv

export OPENAI_API_KEY=sk-proj-...
python client.py

import os
from strands import Agent
from strands.tools.mcp import MCPClient
from strands.models.openai import OpenAIModel
from mcp import stdio_client, StdioServerParameters

# 1) Configure the MCP Client with Stdio transport
# This uses 'uvx' to run the official neo4j-mcp package directly
# Environment variables are passed to configure the specific database connection
def create_mcp_client():
    return stdio_client(
        StdioServerParameters(
            command="neo4j-mcp",
            args=[],
            env={
                "NEO4J_URI": "neo4j+s://demo.neo4jlabs.com:7687",
                "NEO4J_USERNAME": "companies",
                "NEO4J_PASSWORD": "companies",
                "NEO4J_DATABASE": "companies"
            }
        )
    )

def main():
    # 2) Initialize the MCP Client wrapper
    # Strands uses this wrapper to manage the tool discovery and execution
    client = MCPClient(create_mcp_client)

    print("Connecting to Neo4j via MCP...")

    with client:
        # 3) List available tools
        tools = client.list_tools_sync()

        # Helper to print tool names (accessing the internal mcp_tool object)
        tool_names = [t.mcp_tool.name for t in tools]
        print("Available tools:", tool_names)

        # 4) Create a Strands Agent using OpenAI
        # We explicitly use OpenAIModel to avoid defaulting to AWS Bedrock.
        # Note: 'model_id' is required (not 'model').
        # The API Key is picked up automatically from the OPENAI_API_KEY env var.
        agent = Agent(
            tools=tools,
            model=OpenAIModel(
                model_id="gpt-4o"
            ),
            system_prompt="You are a helper for querying graph databases. Use the available tools to answer questions."
        )

        # 5) Ask a question
        query = "What are 5 companies mentioned in articles from January 2023?"
        print(f"\nUser Query: {query}")

        response = agent(query)
        print("\nAgent Response:")
        print(response)

if __name__ == "__main__":
    main()

import os
from strands import Agent
from strands.agent.conversation_manager import SlidingWindowConversationManager
from strands.tools.mcp import MCPClient
from strands.models.openai import OpenAIModel
from mcp import stdio_client, StdioServerParameters

# 1) Configure the MCP Client with Stdio transport
# This uses 'uvx' to run the official neo4j-mcp package
stdio_neo4j_mcp_client = MCPClient(
    lambda: stdio_client(
        StdioServerParameters(
            command="neo4j-mcp"
            args=[],
            env = {
                "NEO4J_URI" :  "neo4j+s://demo.neo4jlabs.com:7687",
                "NEO4J_USERNAME" : "fincen",
                "NEO4J_PASSWORD" :  "fincen",
                "NEO4J_DATABASE" : "fincen"
            }
        )
    )
)

# 2) Create a conversation manager
conversation_manager = SlidingWindowConversationManager(
    window_size=20,  # Maximum number of messages to keep
)

def call_agent(query):
    print('Starting agent query...')

    # 3) Create and run the agent context
    with stdio_neo4j_mcp_client:
        # Fetch available tools dynamically from the MCP server
        tools = stdio_neo4j_mcp_client.list_tools_sync()

        # Initialize the Strands Agent
        agent = Agent(
            tools=tools,
            callback_handler=None,
            conversation_manager=conversation_manager,
            model=OpenAIModel(
                model_id="gpt-4o"
            )
        )

        # Execute the query
        response = agent(query)

    print('Agent response:')
    print(response)
    return response

if __name__ == "__main__":
    # Example Query suitable for the FinCEN dataset
    call_agent("Retrieve the schema of the database to understand the nodes and relationships.")

import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
from neo4j import GraphDatabase

# 1) Direct connection for visualization data retrieval
NEO4J_URI = "neo4j+s://demo.neo4jlabs.com:7687"
NEO4J_USERNAME = "fincen"
NEO4J_PASSWORD = "fincen"

def get_transaction_data():
    with GraphDatabase.driver(NEO4J_URI, auth=(NEO4J_USERNAME, NEO4J_PASSWORD)) as driver:
        # Example query: Count transactions by Entity
        query = """
        MATCH (s:Entity)-[r:FILED]->(f:Filing)
        RETURN s.name as Entity, count(r) as TransactionCount
        ORDER BY TransactionCount DESC
        LIMIT 10
        """
        records, summary, keys = driver.execute_query(query, database_="fincen")
        # Convert to Pandas DataFrame
        return pd.DataFrame([r.data() for r in records])

# 2) Visualize Data
if __name__ == "__main__":
    df = get_transaction_data()

    if not df.empty:
        plt.figure(figsize=(10, 6))
        sns.barplot(data=df, x='TransactionCount', y='Entity', palette='viridis')
        plt.title('Top 10 Entities by Transaction Filings (FinCEN)')
        plt.xlabel('Number of Filings')
        plt.ylabel('Entity Name')
        plt.tight_layout()
        plt.show()
    else:
        print("No data found to visualize.")