Agents are only as good as the context you give them. A knowledge graph provides a memory they can query. This is the smallest setup I could come up with that hands one a real Neo4j graph to explore over the Model Context Protocol (MCP) - one Docker Compose service, a .env for credentials, and a .mcp.json that wires the Neo4j MCP server into Claude Code.

Docker Compose

One container, two ports. The Browser UI on 7474 lets you eyeball the graph; Bolt on 7687 is the wire Claude actually talks over. Persistent volumes mean your graph survives a restart. APOC is Neo4j's add-on toolbox - think of it as a set of extra tools that don't come with the database out of the box but that nearly everyone ends up needing, so we switch it on from the start.

docker-compose.yaml
services:
  neo4j:
    image: neo4j:5
    ports:
      - "7474:7474" # Browser UI
      - "7687:7687" # Bolt protocol
    volumes:
      - neo4j_data:/data
      - neo4j_logs:/logs
    environment:
      - NEO4J_AUTH=${NEO4J_AUTH:-neo4j/changeme}
      - NEO4J_server_memory_heap_initial__size=512m
      - NEO4J_server_memory_heap_max__size=1G
      - NEO4J_server_memory_pagecache_size=512m
      - NEO4J_PLUGINS=["apoc"]
    restart: unless-stopped


volumes:
  neo4j_data:
  neo4j_logs:

Environment file

Neo4j wants credentials as a single username/password string, not two fields. Easy to trip on. Keep it in .env and out of git.

Terminal window
# Neo4j credentials (format: username/password)
NEO4J_AUTH=neo4j/changeme

Install the MCP server

No global install needed. uvx runs the Neo4j MCP server straight from PyPI in an ephemeral environment - first run downloads it, later runs reuse the cache. The package is neo4j-mcp-server; its CLI is neo4j-mcp, so point uvx at the package with --from.

Terminal window
# one-off check that it launches
uvx --from neo4j-mcp-server neo4j-mcp

MCP configuration

This is the wiring. .mcp.json tells Claude Code to spawn the Neo4j MCP server via uvx, which turns Claude's requests into Cypher and runs them against the graph over Bolt. Change changeme to match your .env, or nothing connects.

.mcp.json
{
  "mcpServers": {
    "neo4j": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "neo4j-mcp-server", "neo4j-mcp"],
      "env": {
        "NEO4J_URI": "bolt://localhost:7687",
        "NEO4J_USERNAME": "neo4j",
        "NEO4J_PASSWORD": "changeme",
        "NEO4J_DATABASE": "neo4j"
      }
    }
  }
}

Run it

  1. Run docker compose up -d to start Neo4j.
  2. Open the Browser UI at http://localhost:7474.
  3. Seed a few nodes to populate the graph.
  4. Launch Claude Code in the project directory and query the graph via MCP.

That's the whole thing. No cloud account, no API keys to rotate, no schema migrations. Just a graph database on localhost that Claude can walk with Cypher - ask it how two nodes connect and watch it write the query itself.