Model Context Protocol (MCP) Integrations for Neo4j

What is the Model Context Protocol

The Model Context Protocol was introduced by Anthropic in Nov 2024 to ease the integration of LLM Assistants with tools providing access to data, services, APIs and local systems.

It comes with a open source specification, and a number of SDKs (Python, TypeScript, Java, etc.) to implement MCP Servers and clients.

The MCP architecture consists of the MCP Host program which can be configured with a number of MCP Servers that provide capabilities like tools, resources, and prompts.

If you want to dive deeper, here is a comprehensive overview with explanations and examples with Neo4j.

Whenever a user (or agent) requests in an interaction some task that could be executed by or through using a MCP Server capability, the LLM will select the tool, extract parameters and invoke the tool through an MCP Client connected to the MCP Server.

mcp architecture neo4j e1743543257487

Large Cloud, Infrastructure and LLM Providers announced support for MCP in their offerings.

Also Agent Frameworks like LangChain, CrewAI, Pydantic.AI and Agent Development Kit (ADK) and others are supporting MCP as tools.

There are still a number of open issues, especially around security, discovery, trusted deployment, permission control in the MCP protocol.

But Antropic and the larger community are determined to address them quickly. Especially with the large providers joining the ecosystem, many of the enterprise related requirements will be addressed soon.

The MCP protocol saw a lot of adoption since the beginning of 2025 with thousands of servers for all kinds of APIs having been built by the community.

MCP Servers

MCP Servers are providing an implementation of the MCP protocol (JSON-RPC) over different transports - standard-I/O or HTTP(S). Most servers today are local servers, with http based servers growing in relevance, esp. with the recent addition of OAuth 2.1 to the MCP specification.

The offer these capabilities:

  • tools - tools can retrieve data or take action, they take parameters and return results or errors to the client

  • resources - read only static or dynamic resources, like files, pages, schemas

  • prompts - multi-step prompts with placeholders to guide user interactions with this service

Some example servers (Neo4j servers explained below)

MCP Clients/Hosts

An MCP Host (LLM-Application like Claude, Cursor or CoPilot) can be configured with a number of MCP Servers. When an MCP Server is activated, it creates an MCP client for opening a persistent, bi-directional connection to the MCP Server (using JSON RPC).

Then whenever a user requests a task for which a server’s capability can be used, the MCP host LLM:

  1. extracts the parameter and information from the instruction and the conversational context,

  2. selects the best tool based on description and past usage and

  3. sends a request.

  4. the response is then integrated into the current context,

  5. will be used in the conversation and

  6. can be used as input to other MCP calls or

  7. generating artifacts like documents, images, code.

Developer Tools and IDEs

All the AI powered IDEs support MCP now, so that during your development process you can access services from the AI powered chat. The built-in agents can use these MCP servers to fetch information relevant for your project (e.g. database definitions), spin up services or databases or open pull requests or create JIRA tickets.

mcp inspector toolbox

End User Tools

End user tools will be able to use MCP servers to also securely access SaaS services, Document storage (Drive, Sharepoint, OneDrive, Box), generative tools and corporate databases.

copilot studio mcp demo

Neo4j MCP Servers

The first Neo4j MCP server to interact with a Neo4j Database via Cypher was created in December 2024, shortly after MCP was launched.

Since then we have added several more MCP servers, to control the Neo4j Aura API and to store memories as a knowledge graph in Neo4j.

Neo4j also integrated with Google’s MCP Toolbox to provide optimized tool integration for graph databases.

MCP-Neo4j-Cypher

The mcp-neo4j-cypher Server allows to extract the graph database schema so give the agent-LLM to generate Cypher queries to query and update the database.

Tools:

  • get-neo4j-schema

  • read-neo4j-cypher

  • write-neo4j-cypher

1*UnC f LmG5RB wk3G16TgQ
mcp neo4j cypher movies pie

Here is a blog post showing the server in action, to extract information progressively from a database and use it both to drill down more but also render interactive charts and visualizations.

MCP-Neo4j-Memory

MCP came with an example knowledge graph implementation for storing memories, we upgraded that to use Neo4j as a real graph database.

The mcp-neo4j-memory Server stores and updates entities with observations and relationships between them in Neo4j and can search and retrieve subgraphs of relevance.

It also works for long conversations or plans that turn into more comprehensive eknowledge graphs.

It offers these tools and operations:

Querying

  • read_graph load whole graph

  • search_nodes by search string

  • find_nodes by name

Updating

  • create_entities, delete_entities - create memory entities

  • create_relations, delete_relations - manage relationships

  • add_observations, delete_observations - add observation statements to entities

mpc neo4j memory colleagues
mcp memory instructions
mcp memory graph

MCP Neo4j Aura Manager

One interesting area for developers of using MCP servers is actually not the data side but infrastructure. While developing applications, it is useful to spin up or manage database instances, or pause/resume from the comfort of your IDE. Other infrastructure aspects could be migrations, cost overviews, inviting team members etc.

The mcp-neo4j-aura-manager provides access to the Neo4j Aura API for managing databases, tenants, projects.

When configured with client-id and client-secret the MCP server can access the Aura API on your behalf and offers the following tools as operations:

Instance Management

  • list_instances

  • get_instance_details, get_instance_by_name

  • create_instance, delete_instance

  • update_instance_name, update_instance_memory, update_instance_vector_optimization

  • pause_instance, resume_instance

Tenant/Project Management

  • list_tenants

  • get_tenant_details

cursor aura mcp
mcp neo4j aura create instance

Google MCP Toolbox

Google’s MCP Toolbox for Database is an open source tool to integrate a variety of databases as sources for agentic systems.

Individual parameterized queries can be defined as tools and made available to agent frameworks. Toolbox takes care of database connection management, authentication, result processing and error handling.

architecture

Toolbox integrates as an MCP Server but also directly with LangChain and LlamaIndex.

You can deploy MCP toolbox to Cloud Run or Kubernetes or run it locally on your system.

The Neo4j Integration was added early in Dec 2024, and provides

mcp toolbox tools
mcp server toolkit inspector

Learn more in this article: Build AI Agents With Google’s MCP Toolbox and Neo4j Knowledge Graphs - Graph Database & Analytics

MCP with Agent Frameworks

Most Agent Frameworks support MCP servers as tools, and so also the Neo4j MCP servers.

Using MCP Toolbox and Neo4j with the Google Agent Development Kit (ADK)

Google’s new Agent Development Kit (ADK) also supports integrating MCP Servers as tools for agents.

It also natively supports the MCP Toolbox so that tools defined there can be used directly.

This notebook demonstrates examples of using Agent Development Kit with Neo4j.

Using MCP Servers with LangChain / LangGraph

MCP is available in the LangChain ecosystem through MCP Adapters.

LangChain shared also a very interesting discusion on MCP.

Using MCP Servers with LlamaIndex Workflows

MCP Servers are integrated via an MCP Tool Spec into LlamaIndex. Agents can be connected to tools via the integration installed via LlamaHub, as demonstrated in this LlamaIndex MCP Example Notebook

Using MCP Servers with CrewAI

CrewAI provides bi-directional integration, published CrewAI workflows can be exposed as MCP servers.

With work currently ongoing for using MCP servers as crew tools

Using MCP Servers with Pydantic.AI

PydanticAI supports Model Context Protocol (MCP) in three ways:

  • Agents act as an MCP Client, connecting to MCP servers to use their tools

  • Agents can be used within MCP servers

  • As part of PydanticAI, we’re building a number of MCP servers

Using MCP Servers with Semantik Kernel

MCP tools can be converted to to Semantic Kernel functions so they can be added to a Kernel instance.