A CMS MCP server is a small service that gives an AI assistant such as Claude, ChatGPT or Cursor a fixed set of tools for your website content: list pages, read an entry, change a field, publish. You add its URL to the assistant once, sign in, and then ask for edits in plain words. Whether that is safe depends almost entirely on which tools the server offers and what it does when one of them is called.
That makes picking a CMS MCP server less about the chat and more about reading a tool list. This guide sorts the servers you will come across into four kinds, then walks through the questions that separate a careful server from a risky one.
What the Server Actually Is
MCP, the Model Context Protocol, is an open standard for connecting AI applications to outside systems. The assistant you talk to is the host. It connects to one or more servers, asks each one which tools it has, and calls them with arguments. We explain the protocol from the beginning in editing a website from Claude or ChatGPT. This article starts where that one stops: you know what MCP is, and now you have to choose a server for your content.
The key point is that the assistant only ever does what the tools allow. A server with a "read entry" tool and nothing else cannot break anything. A server with a "write any file" tool can do anything a developer could. Everything in between is a design decision made by whoever built the server.
Four Kinds of CMS MCP Server
Servers from headless CMS vendors
Sanity, Contentful, Strapi and Storyblok all document an MCP server for content stored in their platforms. They work on structured entries, and each draws its own limits. Sanity's MCP server saves edits from its patch tool "to the draft or release version, never directly to published content". Storyblok's server keeps deletes in a separate tool that "requires explicit confirmation from the user before use". In Strapi's server, "each MCP session is scoped to the permissions of the token used to connect", and its docs warn that Media Library deletions have "no trash and no undo". Our guide to headless CMS with AI editing compares these vendors in more detail.
If your content already lives in one of these platforms, its own server is usually the right choice.
Servers that expose a traditional CMS
WordPress has an official MCP Adapter that "bridges WordPress's Abilities API with the MCP specification". In practice that means plugins and themes register abilities, and the adapter offers them to AI clients as tools. What an assistant can do then depends on which abilities are installed and switched on, so the tool list can differ from one WordPress site to the next.
General Git hosting servers
If your site is code in a repository, the GitHub MCP server gives assistants "the ability to read repositories and code files, manage issues and PRs". It includes tools to create or update files and to open and merge pull requests. It is a developer tool that sees every file the same way, headline or build setting. It does offer limits. Toolsets switch whole groups of tools on or off, and its docs say "write tools are skipped if --read-only is set". For a developer working in Cursor or Claude Code this is very useful. For a client changing opening hours, it is far more power than the job needs.
Purpose-built editing servers
The fourth kind sits on top of a site's repository or database and offers content-shaped tools instead of raw file access: "read the site map", "change this text", "draft this translation". The rules about what may change live inside the server, so they apply whichever assistant connects. This is the kind most relevant to code-based sites whose owners are not developers.
How to Judge a CMS MCP Server
These questions apply to all four kinds. Most can be answered from the server's documentation before you connect anything.
1. What does the tool list contain?
Ask for the full list and read it. Look for three groups: read tools, write tools and destructive tools (delete, unpublish, overwrite, merge). A good server keeps them apart so you can allow one group without the others. Be wary of a single broad tool such as "execute query" or "write file" when your actual need is changing text.
The MCP specification lets a server label tools with hints, for example that a tool is read-only or may be destructive. Those labels help, but the tools section of the spec is clear that "clients MUST consider tool annotations to be untrusted unless they come from trusted servers". A label is a claim made by the server, not a lock.
2. How does sign-in work, and what does it grant?
For servers reached over the web, the MCP authorization spec says implementations "SHOULD conform" to it, and it is built on OAuth 2.1. In practice you click connect, sign in to the service in a browser window, and the assistant receives a token. The spec also says clients should request "only the scopes necessary for their intended operations".
Two things to check. First, can you limit the connection to one site or one space, rather than every project in the account? Second, can you revoke it in one place and have that take effect straight away? The security best practices list "wildcard or omnibus scopes (*, all, full-access)" as a common mistake, so a server that only offers all-or-nothing access deserves a closer look.
A server that asks you to paste a personal admin token instead is not automatically wrong, especially for local or automated use. But that token is then the only boundary, so make a dedicated one with the fewest permissions the job needs.
3. Who approves a write?
The spec says "there SHOULD always be a human in the loop with the ability to deny tool invocations". Your assistant handles part of that by asking before it calls a tool, and you should leave those prompts on for write tools. The server can do more. It can save to a draft instead of publishing, show a preview and wait for a yes, or turn a risky change into a review request for someone else. Approval on the server side is stronger than a confirmation prompt in the chat window, because it still applies when someone clicks "always allow" without reading.
4. What is protected, whatever the assistant asks?
Some parts of a site should not change through chat at all: layout, navigation, scripts, build files, and often the page title and main heading. Check whether the server lets you mark such areas as locked or "ask first", and whether those rules are enforced in the tools rather than written in a prompt the assistant may ignore. Our article on content editing guardrails covers how to decide which parts of a site belong in which group.
5. Is there a record and a way back?
Every change should be logged with who made it, through which tool, and when. Then check how undo works. A draft can be thrown away. A Git commit can be reverted. A permanent delete cannot be reversed. If the server can delete, find out what deleting really does before you allow it.
6. What does the assistant get to see?
Reading is not harmless. Whatever a read tool returns goes into the conversation with the AI provider. A server that returns a whole database table or every file in a repository shares more than an edit needs. Prefer servers that return the page or entry you asked about, and that keep secrets such as environment files out of reach.
When You Do Not Need a CMS MCP Server
- One-off text changes on a site with a working editor. If your CMS dashboard already works and you change something twice a year, connecting an assistant adds setup and little benefit.
- Design and structure work. New sections, new page types and layout changes are development work. A chat can describe them, but a developer should build and review them.
- Large editorial teams. Scheduled publishing, roles and editorial workflows belong in a CMS built for that. An MCP server can sit next to it, but it does not replace it.
Where CMS Chat Fits
CMS Chat is the fourth kind, for code-based websites in a GitHub repository: Next.js, Astro, Hugo, Nuxt, plain HTML, and sites made with AI builders such as Lovable, Bolt or v0 once they are on GitHub. It is not a WordPress or headless CMS plugin and does not connect to other CMSs. The agency or developer who built the site installs the CMS Chat GitHub App on the chosen repository once, with no content migration.
The CMS Chat MCP server works from Claude, ChatGPT, Claude Code, Codex, Cursor and other MCP clients. You sign in with your CMS Chat account over OAuth 2.1, choose the site the assistant may touch, and can revoke access at any time. The tools follow the same rules as the built-in assistant: small edits to text, prices, photos or opening hours go live as plain Git commits with history and revert; changes to page titles and main headings wait for an OK; and code, layout, navigation and new sections become a request to the agency instead of a silent change. Each change is recorded under the person who made it, with the name of the tool.
Edits from your own assistant are not metered by CMS Chat. Connect AI shows the short setup, and how it works shows the full flow from request to commit.
Frequently Asked Questions
Is a CMS MCP server the same as an AI CMS?
Not quite. An AI CMS is any content system with AI editing built in, and our guide to what an AI CMS is covers the types. An MCP server is one way to reach such a system from an assistant you already use. Some AI CMSs offer one, some do not.
Can I connect more than one CMS MCP server to the same assistant?
Yes. MCP hosts can connect to several servers at once. Keep the names clear and the permissions narrow, so the assistant does not reach for the wrong site or the most powerful tool.
Is it safe to let ChatGPT or Claude publish to my site?
It can be, if the server limits what can change, asks before risky edits, records every change and makes it easy to undo. If the server cannot do those things, keep the connection read-only or limited to drafts.