Overview

CARTO for Agents lets AI agents (Claude, ChatGPT, Cursor, Gemini, and others) work directly with the CARTO platform. Agents can connect to data warehouses, import and export geospatial data, build Builder maps, author and run workflows, expose those workflows as MCP tools, and check usage and activity.

There are three pieces:

• CARTO CLI. A command-line interface (@carto/carto-cli) that exposes nearly every platform operation as a script-friendly command. Used by humans in a terminal, and by agents through tool use.

• CARTO MCP Server. A hosted Model Context Protocol server that exposes built-in CARTO tools and any workflow you publish, ready to plug into web and desktop AI clients (Claude.ai, ChatGPT, MCP Inspector, Gemini CLI).

• CARTO Agent Skills. A public catalog of skill playbooks at CartoDB/agent-skills that teaches AI coding tools (Claude Code, Codex, Gemini CLI, Skills CLI) how to drive CARTO without re-discovering the API every session.

Choose your tool

You want to…	Use this
Run CARTO operations from a script, CI/CD pipeline, or terminal	CARTO CLI
Let a chat agent (Claude.ai, ChatGPT, Gemini) call CARTO tools mid-conversation	CARTO MCP Server
Have an AI coding agent (Claude Code, Cursor, Codex, Gemini CLI) operate CARTO inside your project	CARTO CLI + CARTO Agent Skills
Expose your own analytical workflow as an MCP tool an agent can call	Workflows as MCP Tools (in CARTO User Manual)

Using all three together

The three pieces are not mutually exclusive. In a coding harness like Claude Code, you can wire up any or all of them depending on what the agent needs to do.

• CARTO CLI runs locally and can be called by any agent that can execute shell commands. The agent uses carto to create maps, execute workflows, import data, manage credentials, and query the warehouse.

• CARTO MCP Server speaks the standard Model Context Protocol, so it works with any compatible MCP client. Confirmed in production: Claude.ai, ChatGPT, Claude Desktop, MCP Inspector, MCP Jam. The agent renders interactive maps inline (via view_map and load_builder_map), calls any workflow you've published as an MCP tool, and discovers your connections and saved maps directly from the conversation.

• CARTO Agent Skills ships installation paths for Claude Code, Skills CLI, Codex, and Gemini CLI today (see Installation). Other coding harnesses can consume the same skills if they support a compatible plugin or rules mechanism. The agent uses the skill to pick the right CLI flags, SQL dialect, async-job patterns, and use-case playbook for each request.

Workflows you build in CARTO can be published as new tools across all three surfaces. The same workflow is reachable from the CLI, from the MCP Server, and referenced by the skill playbooks.

Looking for the AI Agents you build inside Builder maps (in-product agents bound to a map's data and tools)? That's a different product surface. See CARTO User Manual → AI Agents. Looking to embed AI-powered map interaction in your own application? See CARTO for Developers → Build an AI-powered map application.