Model Context Protocol (MCP)

Cube provides a Model Context Protocol (MCP) server through the cube-mcp binary, allowing AI assistants and MCP-compatible clients to interact directly with your NIM environment.

Using the MCP server, AI tools can discover available resources, retrieve data, and execute supported operations through the Cube API.

The current version of cube-mcp exposes 307 auto-generated tools, which may significantly impact context length. A future binary will be released that exposes a curated list of the most commonly used tools to optimize performance and context usage.

Prerequisites

Before starting the MCP server, ensure that:

• cube-mcp is installed
• Cube has been configured with your NIM instance URL
• You have authenticated using either a user account or OpenID client credentials

For installation and authentication instructions, see the quickstart guide.

Starting the MCP Server

Launch the MCP server with:

cube-mcp

The server communicates over standard input and output (stdio), making it compatible with MCP-enabled applications such as AI assistants, IDEs, and automation platforms.

Example Configuration

The following example shows how to configure the Cube MCP server in a compatible MCP client:

{
  "mcpServers": {
    "cube": {
      "command": "cube-mcp"
    }
  }
}

Available Capabilities

Once connected, MCP clients can:

• Discover available NIM resources
• Read resource metadata
• Query entities and records
• Create, update, and delete supported resources
• Access NIM data using the authenticated user's permissions

All operations performed through MCP respect the permissions and access controls configured in NIM.

For advanced use cases, we recommend using the provided skill file: https://github.com/Fastiraz/cube/blob/main/docs/SKILL.md