# Best CMS With an MCP Server in 2026

> Almost every CMS vendor now says it has an MCP server. What an assistant can actually do through it, and under whose permissions, differs more than the announcements suggest.

Buyer Guide by DXP Scorecard Analyst Team · Published 2026-11-15 · Reviewed 2026-11-15 · Ratings updated 2026-11-20
Source: https://www.dxpscorecard.com/editorials/best-cms-with-an-mcp-server

## The short answer

Contentful and SitecoreAI are the best CMS platforms with an MCP server in 2026. Contentful runs a hosted server with OAuth sign-in, scoped to one space and environment, that reads and writes content, content models and assets. SitecoreAI's Marketer MCP lets an assistant build pages, personalization and A/B tests under a certified AI management system. Contentstack, Adobe Experience Manager and Sanity are close behind, and Strapi is the open-source pick.

## Key takeaways

- An MCP server is a second way into your CMS. Your team stops opening the editor for routine work and asks Claude, ChatGPT or Copilot to draft the page, tag the assets or schedule the release. Judge the server the way you would judge a new login screen: who gets in, what they can touch and what gets recorded.
- Having an MCP server is no longer a differentiator. Nearly every platform a marketing team would shortlist has shipped one in the past year or so. The differences are in the details: hosted or installed, sign-in or pasted key, read-only or publish, logged or invisible.
- The safest servers run every action as a named person. When the assistant signs in as you through OAuth, it inherits your role and nothing more. A server that takes one shared API key gives every assistant the same powers and makes the log useless.
- Check what the server cannot do before you plan around it. Several publish landing pages but not blog posts, write entries but not content models, or create but never delete. Those limits are sometimes deliberate safety choices and sometimes just unfinished work.
- Most of these servers are less than a year old. Expect tools to be renamed, permissions to tighten and prices to appear. Ask the vendor how MCP changes are announced and whether the server is covered by your support contract.

## Our picks at a glance

- **Best overall MCP server:** Contentful. Hosted, OAuth sign-in, scoped per environment.
- **Best for marketing operations:** SitecoreAI. Pages, personalization and tests by prompt.
- **Best tool coverage:** Contentstack. About a hundred tools across the suite.
- **Best for Adobe enterprises:** Adobe Experience Manager. One server under existing Adobe permissions.
- **Best for hosted websites:** Webflow. Roles enforced, branches, AI-attributed activity log.
- **Best open-source MCP server:** Strapi. Built into core, free when self-hosted.

## Our analysts' take

A year ago, connecting an AI assistant to a content management system meant a developer, an API key and a weekend. Model Context Protocol (MCP), the open standard that lets an assistant such as Claude or ChatGPT use another product's tools, removed that work. A CMS with an MCP server publishes a list of things an assistant may do, such as find an entry, create a page, upload an image or publish a release, and the assistant does them on request. For a marketing team that means content operations by conversation, from the chat window the team already has open.

The trouble is that the phrase covers very different products. One vendor's MCP server is a hosted service you connect in two clicks, signed in as yourself, with every action in the audit log. Another's is an experimental package a developer installs on a laptop with an administrator key pasted into a file. Both appear as a tick in the same comparison table. Teams that do not look closer discover the difference when an assistant overwrites live content, or when security asks who made a change and the answer is a shared token.

This guide rates ten platforms on the server itself and on the controls around it. We looked at whether the server is official and hosted, how much of the CMS it reaches, whose permissions it runs under and what the platform records about AI actions. Security certifications, built-in AI features and day-to-day upkeep carry weight as well, because an MCP server is only as trustworthy as the platform behind it.

## What to look for

1. **Official, hosted and supported.** Ask whether the vendor builds the server, runs it at a web address and covers it under support. A hosted server connects from Claude or ChatGPT in minutes. One that must be installed on each laptop needs a developer and is skipped by most marketers.
2. **Sign-in as a named person.** The assistant should sign in through OAuth as the person using it and inherit that person's role. Avoid setups where everyone shares one management key, because the assistant then has more access than the person asking.
3. **What it can read, write and publish.** Get the tool list. Check whether it covers the content you work on, whether it reaches content models and assets, and whether publishing and deleting are possible, blocked or gated behind confirmation.
4. **A record of every AI action.** Changes made through the server should appear in the audit log, marked as coming from an assistant and tied to a user. Ask on which plan that log is available, because several vendors reserve it for enterprise contracts.
5. **Controls an admin can set.** Look for read-only modes, per-tool switches, drafts or branches for AI work, and one place to turn the server off. These decide whether you can roll the server out to forty editors or only to two specialists.

## The ratings at a glance

Fit score out of 10, weighted for this topic. Ratings run Limited, Fair, Good, Strong, Excellent.

| Rank | Platform | Fit | AI agent access | Security & compliance | AI for marketers | Low IT upkeep | Marketer self-service | Budget predictability |
|---|---|---|---|---|---|---|---|---|
| 1 | [SitecoreAI](https://www.dxpscorecard.com/platform/sitecore-ai) | 8.3 | Excellent | Strong | Strong | Fair | Strong | Limited |
| 2 | [Contentful](https://www.dxpscorecard.com/platform/contentful) | 8.2 | Excellent | Strong | Strong | Fair | Good | Fair |
| 3 | [Contentstack](https://www.dxpscorecard.com/platform/contentstack) | 8.1 | Strong | Strong | Strong | Good | Good | Good |
| 4 | [Adobe Experience Manager](https://www.dxpscorecard.com/platform/aem) | 8.0 | Excellent | Strong | Strong | Limited | Strong | Limited |
| 5 | [Sanity](https://www.dxpscorecard.com/platform/sanity) | 7.9 | Excellent | Good | Good | Fair | Good | Good |
| 6 | [Optimizely SaaS CMS](https://www.dxpscorecard.com/platform/optimizely-saas) | 7.7 | Strong | Good | Strong | Good | Good | Limited |
| 7 | [Webflow](https://www.dxpscorecard.com/platform/webflow) | 7.4 | Strong | Good | Good | Fair | Strong | Fair |
| 8 | [HubSpot Content Hub](https://www.dxpscorecard.com/platform/hubspot-cms) | 7.1 | Good | Good | Good | Fair | Strong | Good |
| 9 | [Strapi](https://www.dxpscorecard.com/platform/strapi) | 6.2 | Good | Good | Limited | Fair | Fair | Good |
| 10 | [WordPress](https://www.dxpscorecard.com/platform/wordpress) | 5.9 | Good | Fair | Fair | Fair | Good | Strong |

## The shortlist, reviewed

### 1. SitecoreAI (Fit 8.3/10)

The server built for marketers instead of developers. The Marketer MCP creates pages, adds and moves components, sets up personalization and A/B tests and works on campaign briefs, with safeguards around running experiments and confirmation before deletes. Sitecore verifies it with Claude, Copilot Studio and Cursor and holds ISO 42001 certification for AI management.

- **Best for:** Enterprise marketing teams that already run SitecoreAI and want campaign and page work done from an assistant under existing workflow and approvals.
- **Watch out for:** It is oriented to pages and campaigns. Defining content types and authoring arbitrary items still go through other APIs, there is almost no bring-your-own-model option, and AI use has a fair-use cap of 50,000 interactions a year.
- **What it costs:** Quote-only. No published tiers or free plan.
- **Full scorecard:** https://www.dxpscorecard.com/platform/sitecore-ai (Markdown: https://www.dxpscorecard.com/platform/sitecore-ai.md)

### 2. Contentful (Fit 8.2/10)

The most complete package for a team that wants to connect and go. The hosted Remote MCP server has been generally available since July 2026, signs in with OAuth and is scoped to one space and environment. It reads and writes entries, content models and assets and can run Contentful's AI Actions from Claude Code, ChatGPT, Cursor and VS Code.

- **Best for:** Enterprise and mid-market content teams that want assistants working in a governed content platform, with an audit log that tells AI Actions apart from human edits.
- **Watch out for:** Scoping stops at the environment and your role, with no documented field-level limits for agents. The older local server had a high-severity token advisory in August 2026, since fixed, so use the hosted one. Log streaming is an enterprise feature.
- **What it costs:** Free plan, Lite at $300 per month; Premium and Enterprise quote-only. AI Actions are metered.
- **Full scorecard:** https://www.dxpscorecard.com/platform/contentful (Markdown: https://www.dxpscorecard.com/platform/contentful.md)

### 3. Contentstack (Fit 8.1/10)

The widest tool catalog here. The official server exposes about a hundred content management tools, including audit logs, version history and workflow, plus delivery, analytics, Brand Kit, Launch hosting, automations and personalization. One OAuth login covers most of it and the assistant inherits that user's permissions.

- **Best for:** Composable enterprise teams with developers on hand that want an assistant to reach the whole product suite, not only entries.
- **Watch out for:** It still runs locally: each user starts it from a command line, and the hosted, profile-scoped edition with per-call audit logging has not shipped. That keeps it a tool for technical users today.
- **What it costs:** Free plan, Build at $29 per month, Growth at $299 per month plus $25 per extra user; Enterprise quote-only. AI runs on metered credits.
- **Full scorecard:** https://www.dxpscorecard.com/platform/contentstack (Markdown: https://www.dxpscorecard.com/platform/contentstack.md)

### 4. Adobe Experience Manager (Fit 8.0/10)

One server for the whole platform. Adobe's unified MCP server handles pages, content fragments and assets, runs governance checks and manages cloud environments and pipelines. Users sign in with their Adobe ID and tools run under their existing permissions, with separate read-only and read-write versions. It also powers Adobe's official Claude connector and ChatGPT plugin.

- **Best for:** Large enterprises already on Adobe Experience Manager as a Cloud Service that want assistants working inside established permissions and brand governance.
- **Watch out for:** Cloud Service only, with minimum release requirements, so self-hosted and older installations are left out. Brand governance for agents comes with a separately sold Agentic package, and this remains the most expensive platform in the guide.
- **What it costs:** Quote-only. Agents are an additional Agentic SKU. Implementations commonly run $500,000 to more than $1 million in services.
- **Full scorecard:** https://www.dxpscorecard.com/platform/aem (Markdown: https://www.dxpscorecard.com/platform/aem.md)

### 5. Sanity (Fit 7.9/10)

The server developers praise and marketers can still use. Sanity hosts it, accepts OAuth or a token, and one command connects Claude Code, Cursor or VS Code. It manages documents and the media library, lands edits as drafts and can check a write against your content model first. A second, read-only server exists for agents that should only look.

- **Best for:** Teams with a developer or an agency partner that want structured content an assistant can safely query, draft and update.
- **Watch out for:** The validation tool is still in beta, tools that spend AI credits cannot be permissioned one by one, and audit logs sit on the Enterprise plan. Custom instructions guide the built-in agent but do not bind it.
- **What it costs:** Free plan, Growth at $15 per seat per month, Enterprise quote-only. AI credits are published at $0.05 each.
- **Full scorecard:** https://www.dxpscorecard.com/platform/sanity (Markdown: https://www.dxpscorecard.com/platform/sanity.md)

### 6. Optimizely SaaS CMS (Fit 7.7/10)

A young server on a platform with mature AI controls. Since June 2026 the CMS server lets Claude and other clients model content, generate components and run CMS operations as the signed-in user. Separate servers cover experimentation and analytics, and admins can write guardrail policies that apply to agents across the product.

- **Best for:** Enterprises that buy Optimizely for content and experimentation together and want one assistant to work across both.
- **Watch out for:** The CMS server is months old and its tool set and documentation are thinner than the longest-running servers. Everything is quote-only and AI runs on metered credits.
- **What it costs:** Quote-only. Priced on traffic and feature tier, with experimentation and AI as add-ons on credits.
- **Full scorecard:** https://www.dxpscorecard.com/platform/optimizely-saas (Markdown: https://www.dxpscorecard.com/platform/optimizely-saas.md)

### 7. Webflow (Fit 7.4/10)

The best server for a website team without developers. Version 2.0, released in July 2026, lets an assistant build pages, collections and components against your design system. It enforces roles down to the site, collection, page and locale, stages work on a separate branch and marks every action as AI in the activity log.

- **Best for:** Design-led marketing teams that want an assistant to build and update a hosted website they then manage visually.
- **Watch out for:** The tools do not yet cover everything in the product, the content model stays shallow next to a headless CMS, agent-driven builds consume metered credits and you cannot bring your own model for Webflow's built-in AI.
- **What it costs:** Published. Site plans from $15 per month, Premium at $25 per month (annual), workspace seats from $15 to $39 per month.
- **Full scorecard:** https://www.dxpscorecard.com/platform/webflow (Markdown: https://www.dxpscorecard.com/platform/webflow.md)

### 8. HubSpot Content Hub (Fit 7.1/10)

The server that reaches the customer database as well as the content. HubSpot's hosted MCP server signs in with OAuth and respects user permissions. Since autumn 2026 it creates and updates landing pages, schedules and sends marketing email and builds campaigns, with every write attributed in the audit log.

- **Best for:** Marketing teams that run campaigns, email and landing pages in HubSpot and want one assistant across all three.
- **Watch out for:** Blog posts and website pages are read and report only through the server; an assistant cannot write them yet. Templates use HubSpot's own language, and the step from Starter to Professional is steep.
- **What it costs:** Published. Starter from $15 per month, Professional from $450 per month (three seats), Enterprise from $1,500 per month. AI runs on HubSpot Credits.
- **Full scorecard:** https://www.dxpscorecard.com/platform/hubspot-cms (Markdown: https://www.dxpscorecard.com/platform/hubspot-cms.md)

### 9. Strapi (Fit 6.2/10)

The open-source option with the server in the core product. Generally available since June 2026 and free when self-hosted, it builds its tools from your content types, covers drafts, publishing and the media library, and re-checks the user's permissions on every request. It is off until an administrator enables it.

- **Best for:** Teams with developers that want a self-hosted or Strapi Cloud backend and full control over which model and assistant connect.
- **Watch out for:** Seeing assistant changes in the audit log needs an Enterprise license. There is no approval gate designed for AI writes, built-in AI features are thin, and self-hosting leaves upgrades and security with your team.
- **What it costs:** Free open-source core. Strapi Cloud from $35 per project per month, Pro at $90 per month; the Growth license for AI features is $45 per month. Enterprise is quote-only.
- **Full scorecard:** https://www.dxpscorecard.com/platform/strapi (Markdown: https://www.dxpscorecard.com/platform/strapi.md)

### 10. WordPress (Fit 5.9/10)

The baseline most teams already own. The official MCP Adapter plugin, maintained by the WordPress core AI team, exposes only the abilities an administrator switches on and checks each user's permissions. The official AI plugin installs it in one step and adds a screen for choosing which abilities assistants may use.

- **Best for:** Teams already on WordPress that want to try assistant-driven work without changing platform or paying for a license.
- **Watch out for:** The adapter is a pre-1.0 plugin outside core, and write abilities beyond an experimental set are still being built. Request logging and approvals are opt-in experiments, and hosting, patching and security remain yours.
- **What it costs:** Free software. Budget for managed hosting from roughly $25 to $35 per month and premium plugins.
- **Full scorecard:** https://www.dxpscorecard.com/platform/wordpress (Markdown: https://www.dxpscorecard.com/platform/wordpress.md)

## Three kinds of MCP server hide behind the same label

When a vendor says it has an MCP server, find out which of three things it means. The answer decides who on your team can use it, how much risk you take on and whether security will approve it at all.

- **Hosted, with personal sign-in.** The vendor runs the server at a web address and each person signs in through OAuth, so the assistant works with that person's role. Contentful, SitecoreAI, Adobe Experience Manager, Sanity, Webflow and HubSpot work this way. It is the only kind a non-technical marketer can connect alone, and the easiest for security to approve.
- **Installed and run by you.** The server is official, but it runs on a laptop or inside your own installation. Contentstack's starts from a command line, Strapi's is switched on in the server configuration and WordPress uses a plugin. These can be just as capable, and self-hosting keeps data in your hands, but someone technical has to set them up and keep them current.
- **Experimental or unofficial.** Some servers carry a labs or beta label and may change without notice, and others are community projects the vendor does not support. Storyblok still describes its hosted server as an experimental labs project, for example. These are fine for a pilot. Do not build a team process on one until the vendor commits to it.

## Questions to ask before you connect an assistant

- **Whose permissions does it use?** Ask whether the assistant acts as the signed-in person or through a shared key. Then test it: sign in as an editor who cannot publish and ask the assistant to publish. Webflow, Strapi and Adobe Experience Manager all document that the user's own permissions apply to every action.
- **Can it publish or delete?** Find out which actions are possible, which are blocked and which need confirmation. SitecoreAI puts confirmation steps in front of deletes, Sanity lands edits as drafts and Webflow stages work on a branch. If nothing stands between a prompt and your live site, add a rule of your own before rollout.
- **Where do AI actions show up?** Ask to see an assistant's change in the audit log. Contentful's log tells AI Actions apart from human edits, Webflow and HubSpot attribute each write, and Strapi marks entries that arrived through the server. Then ask which plan includes that log.
- **What does it not cover?** Read the tool list against your real weekly work. HubSpot's server writes landing pages but not blog posts, and SitecoreAI's handles pages and campaigns but not content types. A gap in the one task your team does most matters more than a long list elsewhere.
- **What will it cost next year?** Server access is rarely priced on its own today, but the AI features they trigger often run on metered credits, and some governance tools are sold separately. Ask for the credit rate, the caps you can set and a written answer on whether server access stays in your plan.

## Frequently asked questions

### What is an MCP server in a CMS?

MCP, the Model Context Protocol, is an open standard that lets AI assistants such as Claude, ChatGPT and Copilot use another product's tools. A CMS with an MCP server lets an assistant find, create, update and sometimes publish content on request, using the permissions of the person signed in, without a custom integration.

### Which CMS has the best MCP server?

Contentful and SitecoreAI lead in 2026. Contentful offers a hosted server with OAuth sign-in that covers content, content models and assets. SitecoreAI's Marketer MCP covers pages, personalization and A/B tests. Contentstack, Adobe Experience Manager and Sanity are close alternatives, Webflow is the best fit for hosted websites and Strapi is the leading open-source choice.

### Does WordPress have an MCP server?

Yes, through the official MCP Adapter plugin maintained by the WordPress core AI team. It is not part of WordPress core and is still a pre-1.0 release. Administrators choose which site abilities assistants may use, and each request is checked against the user's permissions. WordPress VIP customers get a separate hosted service.

### Is it safe to connect Claude or ChatGPT to our CMS?

It can be, with the right server and settings. Choose one where each person signs in individually, the assistant inherits that person's role, changes land as drafts or need confirmation, and every action appears in the audit log. Start with read-only access or a test environment, and avoid shared administrator keys.

### Do I need a developer to use a CMS MCP server?

Not with a hosted server. Contentful, Sanity, Webflow, HubSpot and Adobe Experience Manager can be connected from an assistant's settings by signing in. Servers that are installed locally or switched on inside your own installation, such as Contentstack's, Strapi's and the WordPress adapter, need someone technical for setup.

## How we rated these platforms

Each rating combines the relevant criteria from the DXP Scorecard's independent scorecards into a five-point scale. The Fit score blends these ratings, weighted for this topic, into one number out of 10, and is recalculated whenever a scorecard is updated:

- AI agent access (55%): Can assistants like Claude and ChatGPT work in our CMS through an official MCP server, under our rules?
- Security & compliance (18%): Will legal, security and procurement sign off?
- AI for marketers (18%): Does built-in AI actually save my team time, with guardrails?
- Low IT upkeep (9%): After launch, how much effort does it take just to keep it running?

Scorecards and platform inclusion are unsponsored: no vendor pays to appear in this guide or to change a rating. Methodology: https://www.dxpscorecard.com/scoring-framework
