MCP Server¶

Access this application template on GitHub

The Model Context Protocol (MCP) provides a standardized interface for AI agents to interact with external systems, tools, and data sources. The MCP Server template gives you a ready-to-use structure for building and deploying an MCP server that hosts your tool logic, resources, and prompts. Agents and other MCP clients connect to this server to discover and execute tools at runtime, without embedding tool code in the agent itself.

Use this template when you want to:

• Centralize tool management in one MCP server deployment that multiple agents or clients (e.g., Agentic Application, Cursor, Claude Desktop) can use.
• Add or modify tools without redeploying your agent; you only update the MCP server.
• Deploy a standalone tool server to DataRobot for production, or run it locally for development.
• Follow the MCP protocol standard so your tools work across MCP-compatible frameworks.

For step-by-step integration of MCP tools with agentic workflows, see Integrate tools using an MCP server.

Key features¶

• Centralized tool management—Define and host all tools in one MCP server; the agent connects as a client and discovers tools at runtime.
• Standardized interface—Tools follow the MCP protocol, making them compatible across different agent frameworks and clients.
• Dynamic tool registration—Add or change tools by updating the MCP server; no need to redeploy the agent.
• Structured tool development—Use the template's decorators and patterns to create custom tools, resources, and prompts with clear docstrings and type hints for LLM use.
• Local and production deployment—Run the MCP server locally for development or deploy it to DataRobot; the Agentic Application template can connect to either.
• DataRobot SDK integration—Use the provided SDK client for DataRobot API access inside your tools.

Use cases¶

Review some use cases that are suited for using the MCP Server template:

• Tool-augmented agentic apps: Build one MCP server that exposes DataRobot or custom tools (e.g., projects, deployments, predictions, external APIs) and connect it to the Agentic Application or other agentic workflows.

• Shared tool back end: Let multiple agents or MCP clients (e.g., Cursor, Claude Desktop, VS Code) use the same set of tools by pointing them at your deployed MCP server.

• Custom tool back end: Implement domain-specific tools (e.g., querying databases, calling internal APIs) in a single deployment and scale the server independently from your agents.

Architecture¶

The template supports the standard MCP architecture: an agent or other client connects to your MCP server over the MCP protocol. The server hosts the logic for tools, resources, and prompts; the client discovers and invokes them without containing tool code. You deploy the MCP server separately from your agent, which allows independent updates and scaling. The Agentic Application template includes built-in MCP client support and can connect to an MCP server deployed from this template (locally or in DataRobot).