Skip to content

Lab_5_MCP_Server

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

Lab_5_MCP_Server

README.md

Lab 5 — Neo4j MCP Server

Connect a Strands Agent to a Neo4j knowledge graph through the Model Context Protocol (MCP). This lab introduces MCP tool discovery, then progresses from pre-written Cypher templates to fully autonomous Text2Cypher agents.

What You'll Learn

  • MCP fundamentals: Agent → MCP Server → Data Source architecture, tool discovery, Streamable HTTP transport
  • Cypher Templates pattern: Pre-written queries in @tool functions for reliable, predictable results
  • Text2Cypher pattern: The agent discovers the graph schema and writes original Cypher from scratch
  • Graph-enriched search: Vector similarity combined with graph traversal for knowledge retrieval

Prerequisites

  • Completed Lab 1 (Neo4j Aura instance with SEC financial data loaded)
  • CONFIG.txt updated with MCP_GATEWAY_URL and MCP_ACCESS_TOKEN
  • AWS credentials configured for Amazon Bedrock access

Note: The MCP server is pre-configured by the lab administrator with full embeddings and indexes. You do not need to complete Lab 4 before starting this lab.

Notebooks

Open each notebook in order. The progression builds from MCP basics to autonomous agents.

Notebook Title What You Build
01_intro_strands_mcp.ipynb Intro to Strands + MCP MCP connection via Streamable HTTP, tool discovery with list_tools_sync(), graph schema inspection
02_graph_enriched_search.ipynb Graph-Enriched Search Cypher Templates — pre-written @tool functions that combine vector search with graph traversal for document context and entity enrichment
03_text2cypher_agent.ipynb Text2Cypher Agent An autonomous agent that discovers the graph schema and writes original Cypher from scratch — the Text2Cypher retrieval pattern

Alternative Frameworks

This lab uses the Strands Agents SDK (AWS-native, built-in MCP support, simpler API). LangGraph is a viable alternative that provides fine-grained control over the agent loop via LangChain MCP adapters — better suited for complex, multi-step workflows.

Sample Queries

Once your agent is running in notebook 03, try these questions about the SEC financial data:

Category Example Question
Exploration "How many companies are in the database?"
Products "What products does Apple offer?"
Ownership "Which asset managers own stakes in NVIDIA?"
Risk "What risk factors does Microsoft face?"
Financials "Show me the financial metrics for Tesla."
Cross-entity "Which companies face risk factors related to cybersecurity?"