Yes, an AI assistant like ChatGPT or Claude can now read and help operate your Adobe Marketo Engage instance through Adobe’s Marketo MCP Server. No, it cannot run Marketo on its own. As of 2026 the server is a limited beta that gives an AI client read and non-destructive access to more than 120 Marketo operations, and a human confirms any change. This piece draws that line precisely, with the current limits and dates.
The headlines say Marketo is now programmable, which is true, and they imply that AI can run your marketing operations for you, which is not. The gap between those two things is exactly what a marketing operations leader needs to understand before wiring an assistant into a live instance. What follows is what the Marketo MCP Server actually does, where it stops, and what a sensible team does with it in 2026.
What the Marketo MCP Server actually is
MCP stands for Model Context Protocol, an open standard that lets an AI client call external tools in a consistent way. Adobe’s Marketo Engage MCP Server is a hosted server that exposes Marketo’s existing REST API to any MCP-compatible AI client. It is best understood as a new front door to the same API, not a new set of powers. As of 2026 it is in limited beta, and access is granted by an allowlist tied to your Munchkin account ID, requested through a sign-up form rather than a settings toggle.
What it can do, in numbers
Adobe’s operations reference lists more than 120 operations across thirteen object groups. The distribution is weighted toward the assets a marketing operations team touches daily.
| Object group | Operations | Object group | Operations |
|---|---|---|---|
| Forms | 24 | Emails | 7 |
| Smart Campaigns | 23 | Leads | 7 |
| Programs | 18 | Folders | 7 |
| Smart Lists | 14 | Static Lists | 7 |
| Snippets | 13 | Tokens | 5 |
| Bulk Export | 5 | Channels and Tags | 4 |
The server is built around read and non-destructive operations. A few delete operations exist, but destructive actions are designed to require explicit human confirmation, so the model does not fire deletions on its own. One current gap worth knowing: the Smart List and Smart Campaign asset operations are not fully wired in yet, and Adobe expects that work to finish around the third quarter of 2026.
Where it runs, and where it does not
Adobe’s documentation names the supported clients: Claude Desktop, Claude Code, Cursor, VS Code with GitHub Copilot, OpenAI Codex, and Glean. The server is model-agnostic, so other MCP-compatible clients work too.
The important detail most write-ups miss: it does not work in the claude.ai web app. The server authenticates through HTTP headers, and claude.ai custom connectors require OAuth, which the server does not offer. Header-based authentication runs fine in the desktop and command-line clients, but the browser version of Claude is out until Adobe adds an OAuth path. If your plan was to connect Marketo to Claude in a browser tab, that specific setup does not work today.
How access and permissions work
The server does not store your credentials. The client passes them on every request through HTTP headers, using either Marketo client credentials (a client ID, client secret, and Munchkin account ID) or an identity bearer token with an organization ID. Effective permissions are inherited from the Marketo API-only user attached to your LaunchPoint service. That inheritance is the practical control: you decide what an AI assistant is allowed to do by scoping that service user. For a first connection, give it read access and nothing more, then widen it deliberately.
What it costs you
The server adds no separate charge and no separate quota. It spends against your instance’s existing daily and concurrent Marketo API limits. This matters more than it sounds, because a single plain-language prompt can fan out into several API calls, so repeated or automated use can consume your quota faster than a person expects. Plan for it: run a dedicated API user, and model the call volume rather than assuming it is free.
Six things it is genuinely good at today
- Instance audits. Ask it to find programs that have not been touched in eighteen months, or assets that no longer route anywhere.
- Drafting smart lists from plain language. Describe a segment in a sentence and have it build the smart list definition for your review.
- Pre-launch program QA. Have it check a program’s smart lists, flow logic, and targeting before you turn it on.
- Lead-routing forensics. Ask why a specific record did not reach MQL, or why it was excluded from a campaign, and have it walk the activity history in plain language.
- Cloning forms at scale. Duplicate and adjust forms across programs without doing each by hand.
- Finding and merging duplicates. Search, identify, and merge duplicate records conversationally, with you confirming the merges.
Adobe and early users have documented these and other patterns in a set of real-world use cases. The common thread is that the assistant does the reading and the drafting, and a person does the deciding.
The four hard limits
- It is a beta, not a finished product. Access is allowlisted and the capability set is still growing. Treat it as early software.
- It is request and response only, not event-driven. Nothing inside Marketo can wake an agent. A form fill, a score change, or a click cannot trigger it. The model has to ask, every time.
- It is not autonomous. It makes no decisions and no predictions. It runs the operations a person directs, and writes and deletes expect human review before they go through.
- It adds no new powers. It inherits every existing API limit and permission. It is a new way to reach the same REST API, not a way around its rules.
Put those four together and the honest summary is that the Marketo MCP Server is an excellent assistant and a poor autopilot. Used as an assistant, with a person in the loop, it is the most practical agent entry point Adobe has shipped.
How this fits Adobe’s bigger 2026 story
At Adobe Summit 2026 the company rebranded Experience Cloud as CX Enterprise and pointed its roadmap at AI agents. For Marketo specifically, the roadmap includes a conversational agent layer and a Journey Agent that turns a campaign brief into a structured journey, with the MCP Server positioned as the way third-party agents connect in. Adobe is layering agents onto the existing platform rather than building a single agent brand, and using open standards so outside tools can call in.
Two clarifications save a lot of confusion. First, Einstein is Salesforce’s AI brand, not Adobe’s. Adobe’s AI brands are Sensei for predictive machine learning and Firefly for generative work. Second, and more important, Adobe’s flagship “AI Assistant” belongs to the Experience Platform stack, and its supported applications are Real-Time CDP, Journey Optimizer, and Customer Journey Analytics. A company that buys Marketo Engage on its own does not get that assistant. The Marketo equivalents are the in-product AI Skills Panel and the Email Designer assistant, which are separate and narrower. Adobe’s keynote material blurs this, so it is worth stating plainly.
If you are wiring agents to Marketo in 2026, know the calendar
Several 2026 platform changes affect anyone building or maintaining Marketo integrations, and they are easy to miss. Support for the SOAP API ended on July 31, 2026. REST calls that pass the access token as a URL query parameter stop working after August 31, 2026, so move them to the Authorization header. The Merge Leads API now accepts at most twenty-five IDs per call. We keep the running detail, alongside how to get help with the cleanup, in our companion piece on Marketo support options for 2026.
Our take
Adopt the Marketo MCP Server now, for reading and assisting, with a scoped API user that starts read-only. It will save real time on audits, QA, and lead-routing questions. Do not sell it internally as automation, because it is not, and framing it that way sets up a disappointment. Revisit the autonomous case when writes and event triggers mature, which the roadmap points toward but has not delivered. This is the same honest posture we bring to every platform: adopt what works today, name the limits, and plan for what is coming without pretending it has already arrived.
They can read your instance and help with operations through the Marketo MCP Server, but they cannot run it on their own. The server gives an AI client read and non-destructive access to more than 120 Marketo operations, and a person confirms any change. It is an assistant, not an autopilot.
Not yet. As of 2026 it is a limited beta with access granted by an allowlist tied to your Munchkin account ID. Treat it as early software: useful, but still growing and gated.
No. The server authenticates through HTTP headers, and the claude.ai web app requires OAuth for custom connectors. It runs in Claude Desktop, Claude Code, Cursor, VS Code with Copilot, and Codex, but not in the browser version of Claude until Adobe adds an OAuth path.
Destructive actions are designed to require explicit human confirmation, so the model does not delete on its own. You also control what is possible by scoping the Marketo API user the server inherits its permissions from. Start that user read-only and widen access deliberately.
It adds no separate charge and no separate quota. It spends against your existing Marketo API limits, and a single prompt can fan out into several calls, so model the volume and run a dedicated API user rather than assuming it is free.
Where to go next
If you are thinking about how AI fits your Marketo operations, our Marketo services page explains how we work, our comparison of Marketo support options in 2026 covers who does the hands-on work, and our note on Adobe-certified Marketo experts covers credentials.
Couch & Associates helps teams set up a least-privilege MCP connection and run an agent-readiness review of their Marketo instance. If you want a straight answer on what an assistant can safely do with your setup, reach out.
