MCP server overview¶

The Model Context Protocol (MCP) is an open standard that lets AI agents, large language models (LLMs), and coding assistants discover and interact with external tools, data sources, and services through a consistent interface. Instead of embedding tool logic inside every agent or client, you deploy an MCP server that hosts tools centrally. Clients connect to that server at runtime to list and invoke tools without custom integration code per client.

DataRobot provides several MCP server options—from a managed platform service to templates you deploy yourself—so you can expose DataRobot capabilities and custom logic to agentic workflows, IDEs, and chat clients.

How MCP works¶

MCP uses a client-server model:

• The client—the reasoning engine or assistant that plans tasks. Examples include a LangGraph agent in the DataRobot Agentic Starter template, or an IDE assistant in Cursor, Claude Desktop, or VS Code.
• The MCP server—a web service that hosts tool logic, resources, and prompts. The server executes tasks such as calling the DataRobot API, running predictions, or querying external systems.
• The protocol—a standard interface for discovery and invocation. Clients ask the server which tools are available and send structured requests to run them.

flowchart LR
    Client["MCP client<br/>(agent, IDE, chat app)"]
    Server["MCP server<br/>(tools, resources, prompts)"]
    DR["DataRobot platform<br/>and external APIs"]

    Client <-->|"MCP protocol"| Server
    Server --> DR 

Why use an MCP server?¶

Using an MCP server to provide tools offers several advantages over embedding tools directly in agent code or deploying each tool separately:

• Centralized tool management—define and host tools in one place that multiple agents or clients can share.
• Standardized interface—tools follow the MCP protocol, so they work across MCP-compatible frameworks and clients.
• Dynamic discovery—clients list available tools at runtime; you can add or change tools on the server without redeploying every client.
• Separation of concerns—scale and update the tool server independently from your agents and applications.

For tool integration patterns inside agentic workflows, see Integrate tools using an MCP server. For connecting IDEs and chat clients, see Connect agentic coding environments to MCP servers.

DataRobot MCP server types¶

DataRobot supports three main MCP server options. The right choice depends on whether you need a managed platform endpoint, a fully customizable server, or an MCP server bundled with an agentic application.

DataRobot Global MCP¶

The DataRobot Global MCP is a persistently deployed MCP server that DataRobot automatically provisions on your instance. Agentic workflows and MCP clients can connect to it using a fixed platform endpoint without deploying a separate MCP application.

Endpoint:

https://{DATAROBOT_URL}/api/v2/genai/globalmcp/mcp 

Key characteristics:

• Managed service—no separate MCP deployment to create or maintain.
• Platform-native access—connect from agentic workflows or coding environments using your DataRobot API key.
• Predictive AI tools (initial release)—the Global MCP currently exposes DataRobot predictive AI tools; additional tool categories are planned for future releases.

Authentication: Requests require a DataRobot API key passed as a Bearer token. Obtain your key from the user menu under API keys and tools. See API key management.

When to use it:

• You want immediate access to DataRobot predictive tools from Cursor, Claude Desktop, VS Code, or an agentic workflow.
• You do not need custom tools, third-party integrations, or deployment-specific configuration.
• Your environment provides the Global MCP endpoint (available in recent DataRobot releases).

For client configuration examples, see Using the DataRobot Global MCP.

Standalone MCP server¶

The DataRobot MCP template (af-component-datarobot-mcp) is an App Framework component that deploys a FastMCP-based MCP server as a DataRobot custom model application. You can run it locally for development or deploy it to DataRobot for production.

Endpoints:

Key characteristics:

• Comprehensive built-in tools—pre-built tools for DataRobot platform operations, third-party integrations, and web search.
• Dynamic tool registration—automatically expose tagged DataRobot deployments as MCP tools.
• Custom tool authoring—add domain-specific tools using FastMCP decorators and Python type hints.
• Repeatable instances—apply the component multiple times in one project under different names for separate MCP backends.
• OpenTelemetry tracing—optional observability for tool calls in production.

Built-in tool categories include:

• DataRobot platform—catalog and datasets, modeling and projects, deployments, batch and real-time predictions, vector databases, use cases, and documentation lookup.
• Data connectors—Confluence, Jira, Google Drive, and Microsoft 365 (OAuth configuration required).
• Web search—Perplexity and Tavily (API keys required).

When to use it:

• You need custom tools, integration with collaboration platforms, or full control over the tool surface.
• Multiple agents or MCP clients should share one tool backend.
• You want to turn DataRobot deployments into tools automatically via dynamic tool registration.

For deployment and development, see the MCP Server application template and the MCP template repository.

Agentic Starter MCP server¶

The DataRobot Agentic Starter template includes an MCP server alongside its FastAPI backend, React frontend, and LangGraph agent workflows. The server can run locally during development or deploy to DataRobot with the rest of the application.

Endpoints:

Set MCP_SERVER_PORT in your .env file to change the local port.

Key characteristics:

• Integrated agent client—LangGraph workflows receive MCP tools at runtime through mcp_tools_context in custompy_adaptor (DRUM) and register.py (DRAgent).
• Co-deployed architecture—agent, API, frontend, and MCP server ship together in one template.
• Composable—the agent can also connect to a standalone MCP server or the Global MCP instead of (or in addition to) the bundled server.

When to use it:

• You are building an agentic application with the Agentic Starter template and want MCP tools available out of the box.
• You prefer a single template that includes both the agent and its tool server for local development and deployment.

The bundled MCP server follows the same MCP protocol as the standalone template. For agent-side integration details, see Integrate tools using an MCP server.

エンドポイントリファレンス¶

Use the following patterns when configuring agents or MCP clients. Always use the exact URL from your deployment output when available.

認証¶

Remote MCP connections to DataRobot typically require:

• Authorization: Bearer <DATAROBOT_API_TOKEN>—DataRobot API key for authentication.
• x-datarobot-api-token: <DATAROBOT_API_TOKEN>—required for tool execution on many standalone and Agentic Starter deployments.

The Global MCP generally requires the Bearer token only. See Connect agentic coding environments to MCP servers for per-client configuration examples.

Choose an MCP server¶

Use this guidance to select the right option:

• Start with the Global MCP if you need quick access to DataRobot predictive tools from an IDE or agent and your instance provides the endpoint.
• Use the standalone MCP template if you need custom tools, third-party integrations (Jira, Confluence, Google Drive, Microsoft 365), web search, dynamic deployment registration, or a dedicated tool server shared across multiple applications.
• Use the Agentic Starter MCP server if you are building on the Agentic Starter template and want MCP tools integrated with minimal setup alongside your LangGraph agent.

You can combine options: for example, an Agentic Starter agent can connect to its bundled MCP server for development and to a production standalone MCP server or the Global MCP in deployed environments.

MCP vs. direct tool integration¶

DataRobot also supports integrating tools directly into agents without an MCP server (for example, via ToolClient and direct tool deployments). MCP is preferable when you want centralized tool management, protocol-standard compatibility, and runtime tool discovery.

For direct tool integration, see Add tools to agents.

次のステップ¶