AI assistants like Claude and ChatGPT can write text, but out of the box they cannot touch your WordPress site. An MCP server is the piece that changes that: it gives an AI agent a documented, structured way to read and act on a real site. This guide explains what an MCP server for WordPress is, how it works, which types exist, and what an agent can actually do through one.
What is an MCP server for WordPress?
An MCP server for WordPress is software that exposes the capabilities of a WordPress site (reading and writing posts, managing plugins, querying the database) as structured tools that an AI agent can call through the Model Context Protocol. The AI client decides what to do; the MCP server is the interface that lets it do it on your site.
MCP, the Model Context Protocol, is an open standard introduced by Anthropic in 2024 and now supported by every major AI client. It defines a common language between an AI application (the client) and an external system (the server). Instead of pasting content back and forth or writing custom API code, the client discovers which tools the server offers and calls them directly. For WordPress, that means an agent can list posts, edit a page, activate a plugin or run a query as a sequence of tool calls, each one visible and approvable in the client.
How does MCP work with WordPress?
Three parts are involved, and it helps to keep them distinct:
- The AI client is the application you talk to: Claude Desktop, Claude Code, ChatGPT, Cursor, OpenAI Codex, Gemini CLI and others. It runs the model, holds the conversation and decides which tools to call.
- The MCP server publishes a catalog of tools and executes them when the client calls. For WordPress it translates tool calls into real operations on the site.
- The tools are the individual operations: get a post, update an option, upload a file, run a database query. Each tool has a name, a description and typed inputs, which is what lets the model use them reliably.
The connection between client and site is authenticated. With Novamira, for example, every request needs a WordPress Application Password over HTTPS or an OAuth authorization, and only administrator users can connect. The data path is direct: the client talks to your site, with no third-party server in between.
What types of MCP servers for WordPress are there?
The category covers quite different architectures. Most servers you will meet fall into one of three types.
Remote servers built on the REST API
These run outside WordPress, often as a hosted service, and translate MCP tool calls into WordPress REST API requests. They are quick to set up and need little or nothing installed on the site. The trade-off is that they can only do what the REST API exposes: standard content operations work well, while anything the REST API does not cover (code, files, most plugin internals) stays out of reach, and your content passes through the provider’s infrastructure.
The WordPress core MCP Adapter
Since WordPress 7.0, core itself ships AI building blocks: the Abilities API, which lets WordPress and plugins register what they can do in a machine-readable form, and the MCP Adapter, which exposes those abilities over MCP. This is the foundation the platform is standardizing on. On its own it provides the abilities that core and your installed plugins register, which today is a baseline rather than a complete toolset.
Plugin servers that run inside WordPress
These install as a plugin and run inside the WordPress process itself. Because they live where the data lives, they can go beyond the REST API: direct database queries, file management, PHP execution, WP-CLI style operations, and deep integration with page builders such as Elementor and Bricks. Novamira belongs to this type: it builds on the core Abilities API and MCP Adapter and adds a full set of abilities on top, from content and media to the database schema and code. The connection stays direct between your AI client and your site.
A fourth pattern worth knowing is the CLI bridge: a command-line tool, such as the Novamira CLI, that connects a terminal-based agent to one or more sites without adding a remote server in the middle.
What can an AI agent do through an MCP server?
With a capable server connected, an agent can work on a site the way a developer would, one operation at a time:
- Create and edit posts, pages and custom post types, including custom fields.
- Build and modify page builder layouts that stay editable in Elementor and Bricks.
- Inspect and change plugins, themes, menus, users and options.
- Read the database schema and run queries to answer questions about the data.
- Write and run code to implement a feature or fix a bug.
What a given agent can do in practice depends on the server’s catalog of tools. The full list of what Novamira exposes is on the tools page. For production sites, Novamira Pro includes Read Mode, which offers read abilities only, so an agent can inspect a site without changing it.
Do you need a plugin now that WordPress core speaks MCP?
The core MCP Adapter and a dedicated server are not alternatives; they are layers. Core provides the protocol plumbing and the abilities that core and plugins register. A server like Novamira uses exactly that foundation and adds what agents need to do real work: a broad ability catalog, authentication flows for the major AI clients, PHP execution with crash recovery, page builder expertise, and skills that encode how your site should be built. If your use case is covered by the abilities already registered on your site, the Adapter alone is enough; the moment you want an agent to build or maintain the site, the added layer is what does the work. The how it works page shows the full picture.
How do you set one up?
For Novamira the short version is: install the plugin from the download page, open your AI client’s connector settings, and authorize with OAuth or an Application Password. The connection guide covers Claude, ChatGPT, Cursor and the other supported clients step by step, including the most common connection errors.
Glossary
MCP (Model Context Protocol): the open standard that defines how AI applications discover and call tools on external systems.
MCP client: the AI application that connects to servers and calls their tools (Claude Desktop, Claude Code, ChatGPT, Cursor, OpenAI Codex, Gemini CLI).
MCP server: the software that exposes a system’s capabilities as tools; for WordPress, it maps tool calls to operations on the site.
Tool: a single callable operation with a name, a description and typed inputs, such as “create post” or “run query”.
Ability: in WordPress terms, a capability registered through the core Abilities API; the MCP Adapter exposes abilities as MCP tools.
Application Password: a per-user credential built into WordPress that authenticates API requests without sharing the account password.