What is MCP for a CMS?
The Model Context Protocol gives AI agents a standard way to connect to external tools and data. When a CMS speaks it, an agent such as Claude Code or Cursor can operate the site through its own tools.
The short answer
MCP, the Model Context Protocol, is an open standard that lets an AI agent discover and call tools exposed by a server. A CMS with an MCP server publishes its operations as typed tools. A connected agent can then operate the site directly, with the permissions of the account that authorised it.
What changes when a CMS speaks MCP
MCP is an open protocol introduced by Anthropic in 2024 and now hosted by the Linux Foundation's Agentic AI Foundation. Many clients support it, including Claude, Claude Code, ChatGPT, Cursor, VS Code, Codex. A server declares tools, each with a name, a JSON Schema for its input and usually a description. The agent reads the list, picks a tool and calls it.
A CMS already has a precise vocabulary: pages, fields, collections, media, drafts. An MCP server turns it into tools. The agent asks which fields an article has, then writes values into them. It is most precise on structured content: with named fields, it can change the date of an event and nothing else.
MCP, REST API or CLI?
Three doors into the same system, built for three kinds of caller.
| REST API | CLI | MCP server | |
|---|---|---|---|
| Built for | Your code. | A developer in a terminal, and CI. | An AI agent in a conversation. |
| Discovery | You read the documentation, then write the calls. | --help, then you type the command. |
The server lists its tools and schemas; the agent picks. |
| Typical use | Sync with another system. | Push templates, script a migration. | "Add these events and translate them into German." |
With MCP the agent fetches the operations and their input schemas from the server, instead of reconstructing them from documentation.
How Quilium does it
Quilium runs a hosted MCP server that authenticates over OAuth: nothing to install, no token to paste. The address is the same for every client, https://mcp.quilium.io/mcp. In Claude (claude.ai and Claude Desktop), add it as a custom connector under Settings, Connectors. In ChatGPT, create a connector with the same address once developer mode is on (a beta, on paid plans). In Claude Code, it is one line, then run /mcp to sign in:
claude mcp add --transport http quilium https://mcp.quilium.io/mcp
The server exposes typed tools to read the content model, edit pages, fill collections and upload media. It also serves a skills library: guides to Quilium's patterns and good practice, which the agent loads when a task needs them. So the agent gets more than a list of functions: it learns how Quilium expects a collection, a template or a translation to be built. Two public skills, quilium-builder and quilium-webmaster, add the process around the CMS.
The agent acts with your profile on each site, so an editor's agent cannot change the structure. Pages are edited in draft, and the skills tell agents never to publish without your confirmation. Other clients are covered in Connect an agent over MCP, and every tool in the MCP tools reference.
MCP or CLI for a coding agent?
Both, depending on where the agent runs. In a terminal, Claude Code can also drive Quilium through the CLI, which wraps the API: a shell command uses fewer tokens than an MCP tool call, which suits long or repetitive work such as pushing templates or importing content. Agents that run in the cloud, like Claude or ChatGPT on the web, have no terminal. For them, and for quick tasks anywhere, the MCP server is the way in.
Questions people ask about MCP and CMSs.
Is MCP the same thing as an API?
Not quite. MCP is a protocol built for AI agents. An API is an interface your code calls; MCP lets an AI agent discover the available operations and their input schemas by itself, then call them, often on top of the same operations the API exposes.
Is it safe to let an AI agent edit my website?
As safe as the permissions and the review step around it. On Quilium the agent inherits your profile, page changes stay in draft, and publishing waits for a human. Collection entries have no draft, so review the first one.
Does MCP replace the CMS back office?
No. The back office stays the place where people review, correct and publish. MCP adds a second operator for the mechanical work: bulk entry, translation, restructuring. Both act on the same content, through the same rules.
Connect your agent to a real site.
Create a site, add the MCP server to Claude Code or Cursor, and ask for the first page. Building is free, you pay when the site goes live.