Oracle (OCI Generative AI)

Oracle Cloud Infrastructure (OCI) can call CARTO's MCP Server through MCP Calling on the OpenAI-compatible Responses API using Streamable HTTP. MCP Calling is generally available (GA) on OCI Generative AI.

This is a code-based integration and requires an OCI administrator to configure the necessary IAM policies.

OCI Generative AI Agents, the service marketed as Agent Factory, cannot host Agent Factory is not the MCP integration path. OCI Generative AI Agents, the service marketed as Agent Factory, cannot host an MCP server and does not provide an MCP tool type. Its available tool types include RAG, SQL, function calling, HTTP endpoints, and agent tools.

CARTO MCP is connected through OCI Generative AI's Responses API, using MCP Calling.

Authentication

Oracle uses a token-based connection for MCP Calling. OCI passes a pre-obtained bearer token to the MCP server and does not run an interactive OAuth flow.As a result, the CARTO connection uses a CARTO API Access Token.

With the required token configuration, the integration provides access to the read and discovery subset of CARTO capabilities:

• Browsing connections

• Running SQL

• Searching the Data Observatory

For the broader set of CARTO tools, add the SQL, Maps, and LDS APIs to the token.

The following capabilities require OAuth and are therefore not available through this OCI integration:

• Starting a Workflow run

• Map authoring

• Workflow authoring

• Administrative operations

Requirements

Before you start, you need:

• A tenancy in a region that supports agentic features. In EMEA, supported regions include Frankfurt, London, and Riyadh.

• An OCI administrator who can create the required IAM policy.

• A developer comfortable with Python.

• Your CARTO MCP Server URL.

• A CARTO API Access Token with the MCP Server scope.

Install

1. Configure the IAM policy

An OCI administrator must grant your group permission to use the Generative AI service.

The generative-ai-family aggregate resource covers both inference and API-key operations.

Run the following as an administrator:

allow group <your-group> to manage generative-ai-family in compartment <compartment>

See Oracle's IAM policy documentation for more information.

2. Point the OpenAI SDK at Oracle

OCI Generative AI provides an OpenAI-compatible API, so you can use the OpenAI Python SDK by pointing the client at the OCI endpoint:

from openai import OpenAI

client = OpenAI(
    base_url="https://inference.generativeai.<region>.oci.oraclecloud.com/openai/v1",
    api_key="<your-genai-api-key>",
    project="<your-project-OCID>",
)

Replace <region> and <your-project-OCID> with your OCI values.

See Oracle's OpenAI compatibility documentation for more information.

3. Declare CARTO as an MCP tool

Pass the CARTO MCP Server as an MCP tool in the Responses API request:

resp = client.responses.create(
    model="<model>",
    input="What CARTO connections do I have?",
    tools=[{
        "type": "mcp",
        "server_label": "carto",
        "server_url": "https://<region>.api.carto.com/mcp/<account_id>",
        "authorization": "<CARTO API token — raw, without the 'Bearer ' prefix>",
        "require_approval": "never",
        "allowed_tools": ["explore_data", "search_data_observatory"],
    }],
)

See Oracle's MCP Calling documentation for the complete MCP configuration and Responses API syntax.

Two details: the token goes in authorization without the Bearer prefix (OCI adds it), and require_approval controls whether the agent pauses before running a tool — set it deliberately. Use allowed_tools to expose a focused, low-risk set and to suppress view_map, which can't render inline here.