Agentic AI > MCP > Connect agentic coding environments to MCP servers

Connect agentic coding environments to MCP servers¶

You can connect your DataRobot MCP server to standard agentic coding environments—such as Cursor, Claude Desktop, and VS Code—to allow AI assistants in those environments to discover and call your MCP tools, prompts, and resources. This enables you to use DataRobot capabilities (e.g., projects, deployments, predictions, third-party tools) directly from your IDE or chat client.

This guide explains how to configure each client to use an MCP server that you run locally or that is deployed to DataRobot. It applies to any of the following MCP connection options:

An MCP server deployed using the DataRobot Agentic Starter template (datarobot-agent-application).
A standalone MCP server deployed using the DataRobot MCP template (af-component-datarobot-mcp).
An MCP server implemented by the DataRobot Global MCP.

MCP server vs. agent application

This page focuses on connecting MCP clients (Cursor, Claude Desktop, VS Code) to an MCP server. For integrating an MCP server into a DataRobot agentic workflow (e.g., LangGraph agent in the Agentic Starter template), see Integrate tools using an MCP server.

Using the DataRobot Global MCP¶

The DataRobot Global MCP is a service automatically deployed to your DataRobot instance that agentic workflows can use to access tools.

Available tools

The DataRobot Global MCP currently supports tools for predictive AI. This limitation will be removed in a future release.

Configuration¶

The DataRobot Global MCP requires an API key to authenticate requests. You can obtain your API key from the DataRobot UI by opening the user menu and selecting API keys and tools. See API key management for more information.

Once you have your API key, configure the MCP client to use the DataRobot Global MCP endpoint. Refer to the steps that correspond to your MCP client in the tabs below.

CursorClaude DesktopVS Code

.cursor/mcp.json

{
  "mcpServers": {
    "datarobot-mcp": {
      "url": "https://{DATAROBOT_URL}/api/v2/genai/globalmcp/mcp",
      "headers": {
        "Authorization": "Bearer <YOUR_API_KEY>"
      }
    }
  }
}

To verify the connection, save .cursor/mcp.json in the correct location, restart Cursor or reload the window, then in Chat or Composer ask the AI to list or use tools from the DataRobot MCP server.

claude_desktop_config.json

{
  "mcpServers": {
    "datarobot-mcp": {
      "url": "https://{DATAROBOT_URL}/api/v2/genai/globalmcp/mcp",
      "headers": {
        "Authorization": "Bearer <YOUR_API_KEY>"
      }
    }
  }
}

To verify the connection, save claude_desktop_config.json in the correct location, restart Claude Desktop, then ask Claude to use tools from the DataRobot MCP server.

.vscode/mcp.json

{
  "servers": {
    "datarobotMcp": {
      "type": "http",
      "url": "https://{DATAROBOT_URL}/api/v2/genai/globalmcp/mcp",
      "headers": {
        "Authorization": "Bearer <YOUR_API_KEY>"
      }
    }
  }
}

To verify the connection, save .vscode/mcp.json under .vscode/ or your user profile (MCP: Open User Configuration in the Command Palette), reload the window if prompted, then in Copilot Chat ask the AI to list or use tools from the DataRobot MCP server.

Using a standalone MCP server¶

You can also use a standalone MCP server that you deploy to your own infrastructure. To use a standalone MCP server, you need to configure your MCP client to use the standalone MCP server endpoint.

Prerequisites¶

Before configuring your coding environment, ensure that you have:

A running MCP server (local or deployed):
Local
- An MCP server started via the DataRobot Agentic Starter template (default port 9000) or the DataRobot MCP template (default port 8080).
- Base URL (e.g., http://localhost:9000/mcp or http://localhost:8080/mcp).
Deployed
- MCP server deployed to DataRobot; you will need the MCP endpoint URL (e.g., https://{DATAROBOT_URL}/api/v2/deployments/{DEPLOYMENT_ID}/directAccess/mcp).
- Base URL (e.g., https://{DATAROBOT_URL}/api/v2/deployments/{DEPLOYMENT_ID}/directAccess/mcp)
Endpoint and auth
"Authorization": "Bearer <YOUR_API_KEY>" (required for authentication with the MCP server)
"x-datarobot-api-token": "<YOUR_API_KEY>" (required for tool execution)

Finding the deployed MCP endpoint

To find the endpoint of the deployed MCP server:

For the DataRobot MCP template, after deploying run task infra:info or check the Pulumi/output step for MCP_SERVER_MCP_ENDPOINT.
For the Agentic Starter template, the deployment output includes an MCP server endpoint. Use the URL shown there for your client.

Endpoint reference¶

Context	Base URL	Notes
Agentic Starter (local)	`http://localhost:9000/mcp`	Default port is 9000; set `MCP_SERVER_PORT` to change it.
MCP template (local)	`http://localhost:8080/mcp`	Default port is 8080; set `MCP_SERVER_PORT` to change it.
Deployed to DataRobot	`https://{DATAROBOT_URL}/api/v2/deployments/{DEPLOYMENT_ID}/directAccess/mcp`	Use the exact URL from your deployment output.

Configure your environment¶

You can configure your MCP client to use the MCP server endpoint. Refer to the steps that correspond to your MCP client in the tabs below.

CursorClaude DesktopVS Code

Cursor MCP docs

For Cursor's current MCP options, see Cursor's MCP documentation.

Configuration file location:

Project-specific: <project-root>/.cursor/mcp.json
Global: ~/.cursor/mcp.json

For a local MCP server using the DataRobot MCP template on port 8080:

.cursor/mcp.json (local)

{
    "mcpServers": {
        "datarobot-mcp": {
            "url": "http://localhost:8080/mcp",
            "headers": {
                "Authorization": "Bearer <YOUR_API_KEY>",
                "x-datarobot-api-token": "<YOUR_API_KEY>"
            }
        }
    }
}

For a local MCP server using the DataRobot Agentic Starter on port 9000:

.cursor/mcp.json (Agentic Starter local)

{
    "mcpServers": {
        "datarobot-mcp": {
            "url": "http://localhost:9000/mcp",
            "headers": {
                "Authorization": "Bearer <YOUR_API_KEY>",
                "x-datarobot-api-token": "<YOUR_API_KEY>"
            }
        }
    }
}

For a deployed MCP server:

.cursor/mcp.json (deployed)

{
    "mcpServers": {
        "datarobot-mcp": {
            "url": "https://{DATAROBOT_URL}/api/v2/deployments/{DEPLOYMENT_ID}/directAccess/mcp",
            "headers": {
                "Authorization": "Bearer <YOUR_API_KEY>",
                "x-datarobot-api-token": "<YOUR_API_KEY>"
            }
        }
    }
}

Claude Desktop can connect to remote MCP servers over HTTP. Configure the MCP server in claude_desktop_config.json.

Claude Desktop MCP docs

For Claude Desktop's current MCP options, see Claude Desktop's MCP documentation.

Configuration file location:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

For a local MCP server using the DataRobot MCP template on port 8080:

claude_desktop_config.json (local)

{
    "mcpServers": {
        "datarobot-mcp": {
            "url": "http://localhost:8080/mcp",
            "headers": {
                "Authorization": "Bearer <YOUR_API_KEY>",
                "x-datarobot-api-token": "<YOUR_API_KEY>"
            }
        }
    }
}

For a local MCP server using the DataRobot Agentic Starter on port 9000:

claude_desktop_config.json (Agentic Starter local)

{
    "mcpServers": {
        "datarobot-mcp": {
            "url": "http://localhost:9000/mcp",
            "headers": {
                "Authorization": "Bearer <YOUR_API_KEY>",
                "x-datarobot-api-token": "<YOUR_API_KEY>"
            }
        }
    }
}

For a deployed MCP server:

claude_desktop_config.json (deployed)

{
    "mcpServers": {
        "datarobot-mcp": {
            "url": "https://{DATAROBOT_URL}/api/v2/deployments/{DEPLOYMENT_ID}/directAccess/mcp",
            "headers": {
                "Authorization": "Bearer <YOUR_API_KEY>",
                "x-datarobot-api-token": "<YOUR_API_KEY>"
            }
        }
    }
}

If your deployment requires authentication, add the appropriate headers or token mechanism as supported by Claude Desktop and your MCP server (e.g., environment variables or config fields; check Claude Desktop documentation and your server docs).

Debugging Claude Desktop MCP

On macOS, MCP-related logs are often under ~/Library/Logs/Claude/ (e.g., mcp*.log). Use them to troubleshoot connection or auth issues.

VS Code (with GitHub Copilot) includes built-in MCP client support. Configure remote MCP servers in an mcp.json file (HTTP servers use type, url, and optional headers).

VS Code Copilot MCP docs

For VS Code's current MCP options, the configuration schema, and security notes, see Use MCP servers in VS Code and the MCP configuration reference.

Configuration file location:

Workspace: <project-root>/.vscode/mcp.json (appropriate for shared, non-secret settings that you may commit to source control).
User profile: run MCP: Open User Configuration from your coding environment's Command Palette (⇧⌘P / Ctrl+Shift+P). Prefer this location, or another secret-management mechanism, for API keys and other credentials.

Do not commit API keys

If your MCP configuration includes headers such as Authorization or x-datarobot-api-token, do not commit those secrets to source control. Keep credentialed configuration in your user-profile mcp.json or another secure secret store.

You can also use MCP: Add Server in your coding environment's Command Palette for a guided setup. For more options, see Add and manage MCP servers in VS Code.

For a local MCP server using the DataRobot MCP template on port 8080:

.vscode/mcp.json (local, port 8080)

{
    "servers": {
        "datarobotMcp": {
            "type": "http",
            "url": "http://localhost:8080/mcp",
            "headers": {
                "Authorization": "Bearer <YOUR_API_KEY>",
                "x-datarobot-api-token": "<YOUR_API_KEY>"
            }
        }
    }
}

For a local MCP server using the DataRobot Agentic Starter on port 9000:

.vscode/mcp.json (Agentic Starter local)

{
    "servers": {
        "datarobotMcp": {
            "type": "http",
            "url": "http://localhost:9000/mcp",
            "headers": {
                "Authorization": "Bearer <YOUR_API_KEY>",
                "x-datarobot-api-token": "<YOUR_API_KEY>"
            }
        }
    }
}

For a deployed MCP server:

.vscode/mcp.json (deployed)

{
    "servers": {
        "datarobotMcp": {
            "type": "http",
            "url": "https://{DATAROBOT_URL}/api/v2/deployments/{DEPLOYMENT_ID}/directAccess/mcp",
            "headers": {
                "Authorization": "Bearer <YOUR_API_KEY>",
                "x-datarobot-api-token": "<YOUR_API_KEY>"
            }
        }
    }
}

If your deployed server requires a token, configure authentication using HTTP headers supported by your MCP server. Avoid passing tokens in URLs or query parameters to reduce the risk of credential leakage; see Use MCP servers in VS Code and your server docs.

Apply and verify¶

To confirm connectivity:

Save your MCP configuration in the correct location.
Restart your MCP client (or reload the window).
In your MCP client, ask the AI to list or use tools from the DataRobot MCP server.

Troubleshooting¶

Client cannot connect to MCP server¶

Symptoms: The IDE or Claude reports that the MCP server is unavailable, or tools do not appear.

Solutions:

Confirm the MCP server is running:
Local: run curl -i http://localhost:9000/ or curl -i http://localhost:8080/ (adjust the port to match your local MCP server setup). A successful response indicates the server is up.
Deployed: run curl -i <your-mcp-endpoint-url> using the exact URL from your deployment output. If the server requires authentication, you may need to pass a bearer token or other headers; see your deployment documentation. A successful response indicates the server is up.
Check URL and path: Use the exact base URL and path (e.g., /mcp) required by your client and server.
Confirm firewall and network access: For deployed servers, ensure your network allows outbound HTTPS to the DataRobot host.

Tools do not appear in the client¶

Symptoms: Connection seems OK, but the client does not list MCP tools.

Solutions:

Restart the client (Cursor, Claude Desktop, or VS Code) after changing the MCP config.
Check client and MCP server logs for errors (e.g., Cursor: Output → MCP Logs; Claude: ~/Library/Logs/Claude/). A successful connection will show MCP server logs with tool registration and availability information.

Authentication errors¶

Symptoms: Requests to the deployed MCP server return 401 or similar.

Solutions:

Confirm the deployed server's auth requirements (e.g., bearer token).
Ensure the client is configured with the correct token or credentials (environment variables or extension settings).
For DataRobot deployments, ensure DATAROBOT_API_TOKEN (or the equivalent used by the server) is valid and has access to the deployment.

Additional resources¶

Integrate tools using an MCP server — Use an MCP server from a DataRobot agentic workflow (e.g., LangGraph agent).
DataRobot MCP template — Build and deploy standalone MCP servers with DataRobot integration.
DataRobot MCP template — Configure your MCP client — Client setup details from the template repo.
DataRobot Agentic Starter template — Full agentic application including an MCP server.
Model Context Protocol — Official MCP specification.