🔍📊 AgensGraph Data Modeling MCP Server

🌟 Overview

A Model Context Protocol (MCP) server implementation that provides tools for creating, visualizing, and managing AgensGraph graph data models. This server enables you to define nodes, relationships, and properties to design graph database schemas that can be visualized interactively.

🧩 Features

• Data Model Creation: Define nodes, relationships, and properties
• Validation Tools: Ensure structural integrity of data models
• Visualization: Generate Mermaid diagrams for model visualization
• Arrows.app Integration: Import/export models from Arrows visual editor
• Cypher Generation: Create ingest queries and constraint queries
• Example Models: 7 pre-built real-world data models

No database connection. Nothing here reads or writes a graph — the server takes a model as JSON, answers about it, and hands back Cypher for somebody else to run. That somebody is usually the cypher server's write tool, in the same conversation.

🔧 Installation

uvx mcp-agensgraph-data-modeling

# Or from source
git clone https://github.com/skaiworldwide-oss/agensgraph-ai.git
cd agensgraph-ai/mcp-agensgraph/servers/mcp-agensgraph-data-modeling
uv sync

Add it to your claude_desktop_config.json:

"mcpServers": {
  "agensgraph-data-modeling": {
    "command": "uvx",
    "args": [ "mcp-agensgraph-data-modeling" ]
  }
}

CLI flags: --transport (stdio, http, sse; default stdio), --server-host, --server-port, --server-path, --allow-origins, --allowed-hosts, --namespace. Each has an environment variable — AGENSGRAPH_TRANSPORT, AGENSGRAPH_MCP_SERVER_HOST / _PORT / _PATH, AGENSGRAPH_MCP_SERVER_ALLOW_ORIGINS / _ALLOWED_HOSTS, AGENSGRAPH_NAMESPACE. --namespace myapp prefixes every tool name (myapp-validate_node), so several servers can be reached from one client.

📖 Usage

🛠️ Tools

A model is a plain JSON object: {"nodes": [...], "relationships": [...]}. A node carries a label, a key_property and its other properties; a relationship carries a type, a start_node_label, an end_node_label, a key_property and its properties. A property is {"name", "type", "source", "description"}.

Validation — each returns the model with defaults filled in, or names what is wrong.

Interchange and drawing

Cypher generation — none of it is executed here.

The parameter placeholder is the database driver's %(name)s, not $name. Pass the batch as params: {"records": [...]} — a list bound that way arrives as JSONB, which is what UNWIND expects. $records is not rewritten by anything and reaches the server as a syntax error.

get_constraints_cypher_queries answers with a list because a joined script would have to be split on a separator, and a label may hold whatever a label holds — including that separator. Run each item as it stands. Every statement it emits carries IF NOT EXISTS, so applying them again changes nothing.

No uniqueness is asserted on a relationship, even one with a key property: the generated ingest merges on the endpoints and the key, so what the key identifies is one relationship between one pair, while a constraint on the property alone would claim it is unique across the whole label. Asserting that rejects the second relationship to reuse a key between another pair.

Examples

The seven: patient_journey, supply_chain, software_dependency, oil_gas_monitoring, customer_360, fraud_aml, health_insurance_fraud.

📚 Resources

Twelve, all read-only and none of them a database lookup.

💬 Prompt

create_new_data_model — one prompt, which Claude Desktop surfaces as a starter. It takes data_context (what the data is, and what to pay attention to), use_cases (what the model has to answer), and optionally desired_nodes and desired_relationships. Give it sample data alongside the prompt.

📄 License

This MCP server is licensed under the Apache License 2.0, which is what LICENSE, NOTICE and the package metadata declare. You are free to use, modify and distribute it subject to that licence; see LICENSE for the terms.