FalkorDB MCP Server

A Model Context Protocol (MCP) server for FalkorDB, allowing AI models to query and interact with graph databases. FalkorDB MCP Server enables AI assistants like Claude to interact with FalkorDB graph databases using natural language. Query your graph data, create relationships, and manage your knowledge graph - all through conversational AI.

🎯 What is this?

This server implements the Model Context Protocol (MCP), allowing AI models to:

• Query graph databases using OpenCypher (with read-only mode support)
• Create and manage nodes and relationships
• List and explore multiple graphs
• Delete graphs when needed
• Read-only queries for replica instances or to prevent accidental writes

🚀 Quick Start

Prerequisites

• Node.js 18+
• FalkorDB instance (running locally or remotely)
• Claude Desktop app (for AI integration)

Running from npm

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "falkordb": {
      "command": "npx",
      "args": [
        "-y",
        "@falkordb/mcpserver@latest"
      ],
      "env": {
        "FALKORDB_HOST": "localhost",
        "FALKORDB_PORT": "6379",
        "FALKORDB_USERNAME": "",
        "FALKORDB_PASSWORD": ""
      }
    }
  }
}

Running with npx

You can run the server directly from the command line using npx:

Using inline environment variables:

# Run with stdio transport (default)
FALKORDB_HOST=localhost FALKORDB_PORT=6379 npx -y @falkordb/mcpserver

# Run with HTTP transport
MCP_TRANSPORT=http MCP_PORT=3005 FALKORDB_HOST=localhost FALKORDB_PORT=6379 npx -y @falkordb/mcpserver

Using a .env file:

# Using dotenv-cli to load environment variables from .env
npx dotenv-cli -e .env -- npx @falkordb/mcpserver

This is useful for:

• Quick testing and development
• Running the server standalone without Claude Desktop
• Custom integrations and scripting

Docker Compose

Run FalkorDB and the MCP server together:

cp .env.example .env   # create env file; edit to set MCP_API_KEY, FALKORDB_PASSWORD, etc.
docker compose up -d

Note: Skipping the .env file leaves variables like MCP_API_KEY and FALKORDB_PASSWORD empty, which disables API key authentication and uses no database password.

Tip: Set REDIS_ARGS in .env to pass extra flags to the bundled FalkorDB's redis-server, for example REDIS_ARGS=--appendonly yes. The value is split on whitespace, and the auth flags derived from FALKORDB_PASSWORD are appended after it, so they win on conflict.

Flags that write to disk do not yet survive a docker compose down: the image runs redis-server --dir /var/lib/falkordb/data — appended after REDIS_ARGS, so the directory cannot be overridden here — while the falkordb-data volume is mounted at /data. Until #174 moves the mount to the real data directory, RDB and AOF files are written to the container's writable layer.

This starts FalkorDB with health checks and persistent volumes, plus the MCP server pre-configured to connect to it.

The MCP server runs in HTTP transport mode and is exposed on localhost:8080 by default. To connect a client, configure it to use:

• Transport: http
• URL: http://localhost:8080
• API Key: Set via the MCP_API_KEY environment variable (optional)

See docker-compose.yml for the exact port and configuration values.

Installation

1. Clone and install:

   git clone https://github.com/FalkorDB/FalkorDB-MCPServer.git
   cd FalkorDB-MCPServer
   npm install
   

2. Configure environment:

   cp .env.example .env
   

   Edit .env:

   # Environment Configuration
   NODE_ENV=development
   
   # FalkorDB Configuration
   FALKORDB_HOST=localhost
   FALKORDB_PORT=6379
   FALKORDB_USERNAME=    # Optional
   FALKORDB_PASSWORD=    # Optional
   FALKORDB_TLS=false    # Set to 'true' when FalkorDB is behind TLS
   FALKORDB_DEFAULT_READONLY=false  # Set to 'true' for read-only mode (useful for replicas)
   
   # Logging Configuration (optional)
   ENABLE_FILE_LOGGING=false
   

3. Build the project:

   npm run build
   

🤖 Claude Desktop Integration

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "falkordb": {
      "command": "node",
      "args": [
        "/absolute/path/to/falkordb-mcpserver/dist/index.js"
      ]
    }
  }
}

Restart Claude Desktop and you'll see the FalkorDB tools available!

📚 Available MCP Tools

Once connected, you can ask Claude to:

🔍 Query Graphs

"Show me all people who know each other"
"Find the shortest path between two nodes"
"What relationships does John have?"
"Run a read-only query on the replica instance"

Note: The query_graph tool now supports a readOnly parameter to execute queries in read-only mode using GRAPH.RO_QUERY. This is ideal for:

• Running queries on replica instances
• Preventing accidental write operations
• Ensuring data integrity in production environments

There's also a dedicated query_graph_readonly tool that always executes queries in read-only mode.

Parameterized queries: The query_graph and query_graph_readonly tools accept an optional params object so values can be passed separately from the query text (referenced as $name), instead of string-concatenating them into Cypher. This avoids query-injection risks and malformed queries. For example, a query of MATCH (p:Person {name: $name}) RETURN p with params: { "name": "Alice" }. Parameter names (including nested map keys) must be valid identifiers. Note: FalkorDB does not allow parameters in LIMIT/SKIP clauses.

📝 Manage Data

"Create a new person named Alice who knows Bob"
"Add a 'WORKS_AT' relationship between Alice and TechCorp"

📊 Explore Structure

"List all available graphs"
"Show me the schema of the movies graph"
"What properties do Person nodes usually have in the movies graph?"
"What properties are on ACTED_IN relationships in the movies graph?"
"Delete the old_test graph"

Schema discovery: FalkorDB is schemaless, so three tools help an agent orient itself before querying:

• get_graph_schema — returns node labels, relationship types, and (optionally) the connection topology. Each connection is { source, relationship, target } where source and target are arrays of node labels (a node may have multiple labels) and relationship is the relationship type. Topology is derived from a bounded sample of relationships (connectionSampleSize, default 10000) and can be turned off with includeConnections: false on very large graphs.
• get_node_schema / get_relationship_schema — sample up to sampleSize (default 100) nodes/relationships of a given label/type and rank their property keys by frequency, returning the actual sampledCount alongside requestedSampleSize. Useful for spotting property naming drift.

All three schema tools always execute read-only (GRAPH.RO_QUERY), so they are safe to run against replica/read-only deployments.

A typical orientation workflow is: list_graphs → get_graph_schema → get_node_schema / get_relationship_schema → query_graph.

🛠️ Development

Commands

# Development with hot-reload
npm run dev

# Development with TypeScript execution (faster startup)
npm run dev:ts

# Run tests
npm test

# Run tests in watch mode
npm run test:watch

# Run tests with coverage report
npm run test:coverage

# Lint code
npm run lint

# Lint and auto-fix issues
npm run lint:fix

# Build for production
npm run build

# Start production server
npm start

# Inspect MCP server with debugging tools
npm run inspect

# Clean build artifacts
npm run clean

# Full CI pipeline (test, lint, build)
npm run prepublish

Project Structure

src/
├── index.ts                   # MCP server entry point
├── services/                  # Core business logic
│   ├── falkordb.service.ts   # FalkorDB operations
│   └── logger.service.ts     # Logging and MCP notifications
├── mcp/                      # MCP protocol implementations
│   ├── tools.ts             # MCP tool definitions
│   ├── resources.ts         # MCP resource definitions
│   └── prompts.ts           # MCP prompt definitions
├── errors/                   # Error handling framework
│   ├── AppError.ts          # Custom error classes
│   └── ErrorHandler.ts      # Global error handling
├── config/                   # Configuration management
│   └── index.ts             # Environment configuration
├── models/                   # TypeScript type definitions
│   ├── mcp.types.ts         # MCP protocol types
│   └── mcp-client-config.ts # Configuration models
└── utils/                    # Utility functions
    └── connection-parser.ts  # Connection string parsing

🔧 Advanced Configuration

Transport Modes

The server supports two transport modes:

stdio (default)

Used for direct integration with AI clients like Claude Desktop. Communication happens via standard input/output.

MCP_TRANSPORT=stdio

Streamable HTTP

Exposes the MCP server over HTTP for remote or networked access. Supports multiple concurrent sessions via the MCP Streamable HTTP protocol.

MCP_TRANSPORT=http
MCP_PORT=8080
MCP_API_KEY=your-secret-api-key  # Optional but recommended

When using HTTP transport, clients connect by sending a POST request with an initialize message. The server returns an Mcp-Session-Id header that must be included in subsequent requests. API key authentication is enforced via the Authorization: Bearer <key> header when MCP_API_KEY is set.

Testing HTTP transport:

1. Start the server:

   MCP_TRANSPORT=http MCP_PORT=8080 npm start
   

2. Use the MCP Inspector to connect:

   npx @modelcontextprotocol/inspector --transport streamable-http --url http://localhost:8080
   

Note: npm run inspect uses stdio transport. For HTTP, start the server and inspector separately as shown above.

API Key Authentication:

When MCP_API_KEY is set, all HTTP requests must include an Authorization header:

MCP_TRANSPORT=http MCP_API_KEY=my-secret-key npm start

Clients must then send:

Authorization: Bearer my-secret-key

Requests without a valid key receive a 401 Unauthorized response. Auth is only enforced in HTTP mode — stdio mode ignores MCP_API_KEY since only the parent process can communicate.

Using with Docker

**Using pre-bu