Adobe introduced the Marketo Engage Model Context Protocol (MCP) server in April 2026 as part of the CX Enterprise suite at Adobe Summit. This managed, hosted service removes the need for users to build or maintain custom server infrastructure to connect LLMs with marketing automation systems.

The rollout provides AI assistants with secure, direct access to over 100 read and write operations in Marketo. Agents can query, create, and manage key marketing assets, including smart campaigns, program structures, lead records, email templates, forms, and reusable snippets, all through standard natural language workflows. 

This overview explains what the Marketo MCP server is, how it works with the Marketo API, its uses, setup process, and considerations before connecting to a production instance.

What is an MCP server?

Traditionally, integrating external systems with Marketo required structured API connections with defined endpoints, specific payloads, and predictable inputs and outputs. While effective for software, this approach isn’t convenient for users seeking answers to simple questions. 

An MCP server introduces a new interaction model by acting as an intermediary between an AI agent and Marketo. It translates natural-language requests from the AI into appropriate Marketo API calls and returns responses in a format the AI can interpret. This enables the AI to interact with Marketo as a user would, but through code, and to discuss its findings with you.

MCP is most valuable in two areas:

• Enabling less experienced users to query the database using natural language. 
• Allowing any user to access a unified view of records across multiple systems, such as Marketo and Salesforce, without switching platforms.

Along with Salesforce, Klaviyo, and Braze, the MCP server is now available in Marketo. 

The building blocks: MCP plus the Marketo API 

A Marketo MCP server functions as a wrapper for the Marketo REST API. 

Functions like retrieving a lead by email, browsing emails, requesting a campaign, or describing the lead schema, interact directly with the API. Libraries like FastMCP add a minimal layer to these functions, providing brief descriptions for the AI agent about each tool’s purpose and usage. When the AI agent calls a tool, the wrapper forwards the request to the corresponding Marketo function, which then executes the API request.

In summary, the architecture follows this sequence: 

AI agent→MCP tool call→Marketo function→Marketo API→response returned through the chain. 

This generic pattern applies to other platforms with documented APIs, such as Salesforce, Gmail, or Slack. Once you understand this structure, you can extend it beyond Marketo.

MCP server Marketo use cases

So, what can you do with Marketo’s MCP server? A lot of things:

• You can create a personal Marketo assistant by connecting the MCP server to Claude Desktop or a similar tool. The assistant provides a summary of form fills, email engagement, program membership, score changes, and compliance flags, eliminating the need to manually navigate the user interface.

• Cross-system lookups: with a Salesforce or Gmail connector in place, you can pose one question and get an answer that covers both systems—for example, finding out what is happening with a person in Marketo and seeing what their Salesforce lead record indicates. This combined view means that switching between tabs is no longer necessary.

• Custom agents: MCP tool calls can be combined with capabilities such as web search or a knowledge base to create specialized agents. For example, an ICP scoring agent can pull firmographic and engagement data from Marketo and cross-reference it with your ideal customer profile. An advanced duplicate-merge agent can evaluate conflicting field values and make informed decisions. A Slack-based MQL triage bot can answer “why did this lead qualify?” by checking recent activity, then allow a representative to assign the lead and send an outreach email directly from Slack.

• Debugging handoffs: Troubleshooting is a practical and time-saving application. If a lead was not routed to sales as expected, you can ask the assistant to check the relevant handoff fields in Marketo and, if connected, verify the corresponding fields in Salesforce. This approach removes the need to manually retrace the workflow. 

Currently, the tool is not recommended for unsupervised execution of sensitive actions. 

It is better suited to answering questions and helping with well-understood, lower-risk tasks, with human review required for any actions that write to the database.

Marketo MCP server setup 

You can choose between two approaches, depending on your familiarity with coding:

• Plug-and-play approach: Several open templates, built on libraries such as FastMCP, require minimal setup. Simply clone the project, enter your Marketo API credentials as environment secrets, and deploy. Hosting platforms that integrate with GitHub streamline this process, allowing you to sync your code and deploy the server with just a few clicks, eliminating manual configuration.

• Local, hands-on approach: To test changes or add new tools before deployment, you can run the MCP server locally on your laptop. Because AI platforms must access your server over the public internet, you will typically use a tunneling tool to temporarily expose your local port. This method is useful for testing but is not suitable for long-term use. If your laptop sleeps or disconnects, the server and any dependent services will go offline. Additionally, free tunneling tools often generate a new URL each time they restart, requiring you to update connector configurations in Claude or other platforms. For regular team use, deploying to a dedicated host is recommended.

In both cases, you will need a Marketo API user with the appropriate permissions, as well as your standard REST API credentials (client ID, client secret, and endpoint). If you have not yet set up an API-only user in Marketo, this is the first prerequisite to address before proceeding.

Test it before you trust it 

Before deploying an AI agent to your production instance, test the underlying functions directly, separate from the MCP layer. Here’s how:

• Confirm that authentication works and that read operations, such as browsing emails, listing channels, and describing lead fields, perform as expected. 
• Test write operations using dedicated test assets, such as a scratch folder, a disposable smart campaign, or an email program with a future send date. 

This two-step process makes it easier to identify issues if problems arise when integrating with Claude or another AI client.

Include a cleanup step in your process. If your test suite creates assets, ensure it can delete them afterward to prevent clutter in your instance after each test run.

A few things to consider pre-rollout 

Before you roll it out, here are a few important considerations/reminders: 

• Credentials are sensitive. Your Marketo API client ID and secret provide broad access like a login. Treat any URL exposing your MCP server as a password and do not share it publicly. If multiple users access the endpoint, use an authenticated server that requires an API key along with the connection.

• Implement permission scoping. Not all users need write access. Provide separate endpoints or credentials for read-only tasks like reporting. Reserve broader access for higher-risk activities such as campaign management. Avoid granting blanket access. 

• Approval settings are your safety net. Most AI clients let you configure permissions at a granular level. Always allow, require approval, or block. Start by allowing read operations without restriction. Require manual approval for actions that create, update, or delete records. You can relax these controls once you’re more confident about the assistant. 

Conclusion 

The value of an MCP server lies not in new capabilities beyond Marketo’s API, but in eliminating the need to translate plain-language questions into structured system queries. This saves time for marketing operations teams who repeatedly answer similar inquiries, and enables less technical stakeholders to access data independently without relying on platform experts.

This approach is still emerging and is best suited for research and reporting, not fully automating marketing programs. Even so, the ability to turn repetitive lookups into simple conversational interactions can significantly improve how teams access and work with data.