Neo4j MCP Server

A Model Context Protocol server that exposes a Neo4j graph database as tools an AI agent can call. It serves the MCP protocol over HTTP, so it can be reached by any MCP client — including a Databricks Unity Catalog MCP connection.

Tools

Tool	Purpose
`read_neo4j`	Run a read-only Cypher query (`MATCH` / `RETURN`). Write keywords are rejected.
`write_neo4j`	Run a write Cypher query (`CREATE` / `MERGE` / `SET` / `DELETE`).
`get_neo4j_schema`	Return the graph schema — node labels, relationship types, properties, and connection patterns. Call this first so queries use the right labels/relationships instead of guessing.

Requirements

Node.js 20+
A Neo4j database (e.g. Neo4j Aura)

Setup

npm install
cp .env.example .env      # then fill in your Neo4j credentials + a secret token
npm run build
npm start                 # serves on http://localhost:3000

Endpoints:

GET /health — liveness check
POST /mcp — the MCP endpoint (requires Authorization: Bearer <MCP_SECRET_TOKEN>)

Optional: npm run seed loads sample data into Neo4j.

Environment

See .env.example. Copy it to .env and fill in:

Variable	Description
`NEO4J_URI`	Bolt/Aura connection URI, e.g. `neo4j+s://<id>.databases.neo4j.io`
`NEO4J_USERNAME`	Neo4j username
`NEO4J_PASSWORD`	Neo4j password
`PORT`	HTTP port (default `3000`)
`MCP_SECRET_TOKEN`	Bearer token clients must present on `/mcp`

.env is git-ignored. Never commit real credentials.

Exposing it to Databricks (or any remote client)

The server listens on localhost. To let a remote MCP client reach it during development, expose the port with a tunnel:

ngrok http 3000

Then register the public URL as a UC HTTP/MCP connection in Databricks, using Authorization: Bearer <MCP_SECRET_TOKEN>. For production, host the server on a reachable machine instead of a tunnel.

Project layout

neo4j-mcp-server/
├── src/
│   ├── index.ts   # the MCP server (tools + HTTP transport)
│   └── seed.ts    # optional sample-data seeder
├── package.json
├── tsconfig.json
├── .env.example
└── .gitignore