Sitecore MCP: Architecture, Tools, and Real-World Use Cases

I've spent a good part of this year wiring AI assistants into Sitecore projects. The biggest change wasn't the AI itself. It was MCP.

Before MCP, getting an assistant to work with Sitecore meant copying item IDs into a chat, pasting back field values, and hoping nothing got lost along the way. Now the assistant can talk to Sitecore directly. It can read the content tree, create pages, add components, and set up tests, all through a standard protocol.

This post covers how Sitecore MCP is built, what tools it gives you, and where I've actually found it useful.

What Sitecore MCP Actually Is

MCP is an open standard. It lets an AI model connect to an external system and use that system's actions as "tools."

In the Sitecore world, there are two options worth knowing:

• Sitecore Marketer MCP (official): A hosted, remote server from Sitecore that connects AI clients to SitecoreAI. This is the one most new SitecoreAI projects should look at first.
• Community Sitecore MCP server: It is created by users, and is compatible with SitecoreAI (XM Cloud), and the old XM/XP versions. It is used for use in case you have a self-hosted installation or if you need a deeper access.

Both work on the same principle of AI working out what to do and at the MCP working on how to do it in Sitecore.

The Architecture, Layer by Layer

Here's how a request moves through the official Marketer MCP setup.

1. The MCP client

• This is where you type the prompt.
• Supported clients include Claude Desktop, Cursor, VS Code with GitHub Copilot, Copilot Studio, Postman, and the Codex app.
• Any client that supports MCP over Streamable HTTP can connect.

2. The Marketer MCP server

• It's a remote server. Nothing to install or host yourself.
• It reads the prompt's intent and picks the right tool.
• It turns natural language into secure, executable actions.

3. The Agent API

• This is the engine under the MCP server.
• Every MCP tool maps to an Agent API endpoint. A "create page" tool calls the endpoint that creates a page.
• The commands are outcome-based. You don't chain five low-level calls to build a page. You ask for the page.
• It includes rollback support, so you can build an "Undo" option into your own tools.
• You can also call the Agent API directly as a REST API when you don't need an AI in the loop.

4. SitecoreAI

• The action runs in your SitecoreAI tenant.
• The result comes back to the client: item ID, path, route, or whatever the tool returns.

How authentication works

This is the part I was most curious about, and it's handled well.

• The authorization code flow is implemented with the help of Sitecore Identity.
• You log in, give permission to the processes, and identify your organization and tenant.
• Tokens are kept in the tenant context, and each API call is tailored to the tenant's needs.
• Whatever role you have within SitecoreAI sets the boundaries to its behaviour; if you aren’t able to do something via the UI you can’t expect AI to perform the task.

That last point matters. The AI doesn't get special powers. It works with your permissions.

How the community server differs

• It runs locally or in your own environment, via NPM, Docker, or from source.
• It connects to Sitecore through the Item Service API, GraphQL, and Sitecore PowerShell Extensions.
• It offers 100+ tools, including PowerShell, logging, security, and presentation details.
• You configure the endpoints and credentials yourself, so setup takes more effort.

The Tools You Get

The Marketer MCP groups its tools around everyday work in SitecoreAI.

Sites and pages

• Retrieve sites and their page structure
• Create and retrieve pages
• Organize components and layouts on a page

Content and components

• Create, update, delete, and retrieve content items
• Browse available components
• Configure component datasources

Assets

• Search digital assets
• Manage asset metadata

Brand

• Retrieve and manage brand kits
• Browse brand context folders and files
• Create and refine marketing briefs

Testing and personalization

• Create personalized content variants and targeting rules
• Create and update component-level A/B/n tests
• Retrieve flow definitions and configure variants

Worth knowing from the Agent API side

• Add a new language version to an existing page
• Get a page's HTML, screenshot, or preview URL

The screenshot and HTML endpoints are underrated. They let an agent check its own work.

Real-World Use Cases

These are the patterns I keep coming back to.

Spinning up landing pages

• A prompt like "create a landing page for the summer promo" creates the page under the right parent.
• The agent adds components like Hero, Text Banner, and Promo, then fills in content and links.
• You get back the item ID and route, ready to review.
• What used to take 30 minutes of clicking now takes a couple of minutes plus review.

Content scaffolding from designs

• Connect a design tool's MCP alongside Sitecore.
• The agent reads the design and builds matching pages and datasources.
• The community server's maintainers report around 5x faster Figma-to-Sitecore workflows and roughly 70% less manual scaffolding.

Cross-tool workflows

• MCP works across systems, so one prompt can touch several tools.
• Example: pull approved tasks from a project tool, then publish the matching campaign page in Sitecore.
• The AI becomes the glue between your stack, instead of you being the glue.

Localization prep

• Ask the agent to add language versions to a set of pages.
• It handles the repetitive setup so translators can start sooner.

Setting up experiments

• Describe the test in plain language: which component, which variants.
• The agent creates the A/B/n test and configures variants.
• Faster setup means more tests actually get run.

Content audits (community server)

• Use PowerShell and query tools to scan large content trees.
• Find empty fields, broken references, or outdated items.
• Great for legacy XM/XP sites before a migration.

Building AI into Marketplace apps

• You can embed the Marketer MCP inside a custom Sitecore Marketplace app.
• The app's agent discovers available tools on its own.
• When Sitecore adds new tools, your app picks them up without code changes.
• Since MCP is model agnostic, you can switch LLMs later without rebuilding.

Best Practices I'd Recommend

• Make sure to start in the right environment. Understand how the agent does its job before it is allowed to work with real data.
• Use the right privileges. Give the agent the least possible rights it needs to get the job done.
• Remember to check the AI’s outputs and ensure they are correct.
• Be specific with the prompts. Avoid vague instructions like "make some blog page" and use more precise ones.
• Be sure to always use the current URL for the server. Sitecore has relocated Marketer MCP to a new URL and is slowly retiring the old one.
• Use the correct server, whether it is Marketer MCP with Sitecore AI, or community server with XM/XP or deeper access options.

Final Thoughts

Sitecore MCP doesn't replace developers.It eliminates the boring aspects of the job. The structure is simple: client, MCP server, Agent API, SitecoreAI. Authentication is restricted to your tenant and role, so you're in charge. Overall, the tools account for most of your daily activities.

If you have SitecoreAI, link the Marketer MCP to your editor and try one task. It's the quickest way to discover its role in your work.