WordPress MCP: What It Is and How to Connect AI Agents to Your Site

WordPress MCP is the Model Context Protocol applied to WordPress: a typed, permissioned set of tools that an AI client such as Claude, ChatGPT or Cursor can call on your site, so a plain-English instruction becomes a real action instead of a suggestion you still have to perform yourself. WordPress does not ship an MCP server in core. It ships the Abilities API, and a separate layer turns those abilities into MCP tools.

Four different things now answer to the name “WordPress MCP server” and only three of them can reach a site you own: the official MCP Adapter, a self-hosted community server, and hosted MCP built into your platform, such as InstaMCP. The fourth belongs to WordPress.org and exists to submit plugins to the plugin directory.

Choosing between the three turns on one question: whether the AI needs to read your site or change it. Most hosted MCP in 2026 is read-only, so a server that answers questions about a site perfectly well will refuse to publish a post or update a plugin. Our side-by-side of the best WordPress MCP servers covers the full field.

What is a WordPress MCP server?

An MCP server is a program that exposes the capabilities of a system as a structured list of tools an AI model can call. Model Context Protocol is an open standard originally published by Anthropic and since adopted across the major AI vendors.

The problem it solves is boring but real. Language models are good at deciding what should happen and bad at reliably performing it. Before MCP, every AI product wrote its own bespoke integration for every service. MCP replaces that with one contract: the server publishes what it can do, the client discovers it, and the model picks a tool and fills in the arguments.

Three pieces make up any MCP setup:

• MCP client. The AI application you actually talk to. Claude Desktop, Claude Code, ChatGPT, Cursor, Windsurf, VS Code with Copilot and others.
• MCP server. The program that sits in front of the system being controlled and publishes its tools.
• Transport. The connection between them, either a local process on your machine or a remote HTTP endpoint.

The important detail is that a tool is typed. It has a name, a description and a schema of allowed arguments. A model cannot invent a new capability, and it cannot smuggle an arbitrary command through a tool that only accepts a post ID and a title. That constraint is what makes it safe enough to point at a production system.

Now, if we talk about WordPress MCP server, it is an MCP server that exposes WordPress operations as tools. Once connected, an AI client can read and write posts and pages, manage taxonomies and media, inspect plugins and themes, edit metadata and run diagnostics, all from natural language, without anyone opening wp-admin.

In practice that means instructions like these do real work:

• “List every draft post on the client site and tell me which ones have no featured image.”
• “Create a new post titled Holiday Campaign Ideas, set the category to Marketing, and leave it as a draft.”
• “Which plugins on this site have updates available?”
• “Update the SEO title and meta description on these four posts.”
• “Show me the comments awaiting moderation.”

How a WordPress MCP request actually works

1. You type an instruction into your AI client.
2. The client reads the tool list published by the WordPress MCP server and selects the tool that matches, for example a content listing tool.
3. The model fills in the arguments defined by that tool schema, such as post type, status and count.
4. Most clients pause here and ask you to approve the call. You see exactly which tool is about to run.
5. The server checks the token scope, executes against WordPress (usually through the WordPress REST API or a direct internal call), and returns structured data.
6. The model reads the result and either answers you or chains into the next tool.

That approval step in the middle is worth noticing. MCP does not run unsupervised by default. The model picks a tool and you approve it, and the schema limits what that tool could do even if you approve the wrong one.

Does WordPress have an official MCP server?

Yes and no, and the difference decides what you end up installing. WordPress core does not contain an MCP server. Core ships the Abilities API, which is the standard way a plugin or theme declares what it can do. The official bridge from those abilities to MCP is a separate plugin, WordPress/mcp-adapter, maintained under the WordPress GitHub organisation. Separately, WordPress.org runs its own MCP server, and it has nothing to do with your site.

Four things currently answer to the name “WordPress MCP server”. Here is what each one actually controls, and where it lives.

So if the question is “does WordPress have an MCP server for my site”, the answer is that core supplies the abilities and something else has to expose them over MCP.

The three ways to run WordPress MCP in 2026

Three approaches expose a WordPress site over MCP, and they suit very different teams. What separates them is who runs the server, how it authenticates, and how many sites one connection covers.

1. The official path: Abilities API and the MCP Adapter

This is the route the WordPress project itself is building. The Abilities API shipped in WordPress 6.9. The bridge to MCP is a separate plugin, WordPress/mcp-adapter, which is currently on its own release cadence. The WordPress developer blog published the introduction to the adapter in February 2026. The current release is v0.6.1, published 13 August 2026, and it is still pre-1.0: the public API has changed between minor versions, with migration guides for v0.3.0 and v0.5.0. It is also not listed in the WordPress.org plugin directory, so searching Plugins, Add New for “MCP Adapter” finds nothing and there is no one-click update path. Install it with Composer or from a GitHub release.

2. Self-hosted community servers

Several independent MCP servers for WordPress exist, including our own open-source implementation at InstaWP/mcp-wp. These run as a process on your machine or a VM, hold a config file listing your sites, and authenticate with application passwords.

They are genuinely useful if you want full control of the middleware, and they were the only real option before 6.9. The trade-off is that you own the process, the credentials file and the updates. If you are choosing between approaches, our comparison of InstaWP MCP vs WordPress MCP goes deeper.

3. Hosted MCP built into the platform

The third option is to let your host run it. The platform installs the server-side piece, issues the credential and keeps both updated, so you do not install a plugin, write a config file or keep an application password on your laptop. You flip a toggle and paste a URL.

The trade-off is that you get the tool surface your host decided to build, and you get it only on sites you host there. It is also the route where providers differ most, because two hosts can both say they support MCP and mean opposite things by it.

What hosted WordPress MCP gives you, host by host

“Our host supports MCP” covers two very different products. Some hosts ship a read-only bridge so an AI can describe your site. Others let the AI change it. If you are evaluating hosts on MCP, this is the line to check first, because a read-only server cannot run any of the agency workflows further down this page.

InstaMCP: WordPress MCP built into your host

InstaMCP is the MCP server built into InstaWP managed WordPress hosting. Turning it on installs the plugin, generates a scoped token and gives you a connection URL. It is available on a Sandbox plan and above.

What you get once it is on:

Which AI clients can connect

InstaMCP 1.1 supports 13 AI clients: Claude Desktop, Claude Code, Claude.ai, Cursor, Windsurf, Cline, Roo Code, GitHub Copilot, Zed, Gemini CLI, OpenAI Codex CLI, ChatGPT and Continue.dev. Read the InstaMCP 1.1 release notes for what changed.

How to set up WordPress MCP on InstaWP

You need an InstaWP account with a site on a Sandbox plan or above, and an AI client that supports MCP. You do not need Node.js, a config file or an application password. If you have not created a site yet, follow the docs on how to create a WordPress site, or spin up a disposable WordPress sandbox to test on.

Step 1: Enable MCP on the site

Open your InstaWP dashboard, go to the site you want to connect, and find the MCP menu in the left sidebar of the site detail page.

Switch on the Enable MCP toggle.

That single toggle installs the InstaMCP plugin on the site for you. You can confirm it under Plugins in wp-admin, though there is nothing you need to configure there.

Step 2: Copy your connection URL

With MCP enabled, the page shows a connection URL for that site.

Step 3: Connect your AI client

In Claude Desktop, open Settings and go to the Connectors tab.

Choose Add custom connector, paste the connection URL, and give it a name you will recognise later. If you manage several client sites, name them properly now. “Client Site, Acme Corp” is much easier to live with than “WordPress 3”.

Once it is added, click Configure to see every tool the connector exposes.

Watch how to connect WordPress with Claude using WordPress MCP server of InstaWP.

Step 4: Run your first commands

Start with something read-only so you can watch the approval flow before anything changes. Try “List all posts on the WP site”.

Claude asks permission before running the tool, and tells you which tool it wants to use. Approve it and you get structured results back. From there you can move on to writes:

• “Create a draft post about AI trends in WordPress and set the category to News.”
• “Which plugins on this site need updating?”
• “Read the site logs and tell me what caused the last fatal error.”
• “Show me all pending comments and flag anything that looks like spam.”

Connecting ChatGPT instead

The same connection URL works with ChatGPT. Open ChatGPT, go to Settings, then Apps and Connectors, and add a new connector.

Paste the URL, name it, and set authentication to “No authentication” because the token is already carried in the URL. Confirm and create. You now have the same tool surface from ChatGPT.

For a walkthrough of connecting other clients, see our guide on how to connect AI agents to WordPress using MCP.

Account MCP vs Site MCP

This is the part almost no WordPress MCP guide explains, and it is the part that decides whether MCP is workable at agency scale.

Site MCP, which is what we set up above, works inside one WordPress site. Account MCP works outside the sites, across your whole InstaWP account. Think of it as a building manager and an interior designer. The building manager creates units, takes snapshots and upgrades the plumbing. The interior designer rearranges what is inside one flat. Neither can do the other job.

If you are already using InstaWP to manage multiple client sites from one dashboard, account MCP is the AI equivalent of that dashboard.

How to set up the official WordPress MCP Adapter

If you run your own stack and want the core-native route, here is the current process. It replaces the older npm-based instructions that circulated through 2025.

Requirements: WordPress 6.9 or higher, because the adapter depends on the Abilities API in core.

Step 1: Install the MCP Adapter

Install it with Composer, or download a release from the repository and upload it as a normal plugin.

composer require wordpress/mcp-adapter

Step 2: Create an application password

The adapter authenticates with application passwords. Create a dedicated, least-privilege user rather than using your own admin account.

1. In wp-admin, go to Users, then the profile of the user you want to authenticate as.
2. Scroll to Application Passwords.
3. Enter a label such as “MCP Server Access”.
4. Click Add New Application Password.
5. Copy the generated password immediately. WordPress will not show it again.

Repeat this for every site you want to connect. This is the step that hosted MCP removes entirely, and it is also the step most likely to leave credentials sitting in a plain text file on somebody’s laptop.

Step 3: Connect your client

Add the site as an MCP server in your AI client, pointing at the site MCP endpoint and passing the username and application password. The exact config format differs per client, so follow your client documentation.

Test with a read-only prompt first, exactly as above. If the client lists your posts, the connection is good.

How WordPress MCP permissions actually work

The nervous question about MCP is reasonable: you are giving a language model write access to a live website. Here is how the risk is actually contained, using InstaMCP as the worked example.

Everything runs through a typed tool

There is no general-purpose escape hatch. If a tool updates a post, it updates a post. It cannot be talked into deleting a user or running a shell command, because those arguments do not exist in its schema. This is the single biggest difference between a well-built MCP server and handing an agent SSH access.

Scopes ride on the token

InstaMCP uses four scope levels: mcp:read, mcp:write, mcp:delete and mcp:admin. A connection issued at read scope cannot write, no matter how the prompt is phrased. There is also a Safe Mode for when you want to let someone explore without risk.

The powerful tools are off by default

Three tools sit behind a second gate: execute_php, db_query and site_files. They ship disabled. If you call one on a freshly enabled site and get “Forbidden: execute_php is disabled”, that is correct behaviour, not a bug. An administrator has to turn it on deliberately at WP Admin, Settings, InstaMCP, Capabilities.

The account layer works the same way, with independent capability toggles for sites, PHP and WordPress versions, teams, diagnostics, backups and billing, plus a read-only master switch above all of them. Enabling MCP never enables the dangerous parts, because the two gates are controlled separately and in different places.

Practical rules for live sites

The habits that matter most in day to day use:

• Connect at read scope first and widen only when you hit something you genuinely cannot do.
• Use a dedicated least-privilege WordPress user for adapter-based setups, never your own admin login.
• Keep connection URLs and application passwords out of tickets, docs and chat.
• Rotate credentials when someone leaves the team or a laptop changes hands.
• Prove new workflows on staging before you point them at a client site.
• Leave the gated tools off unless a specific job needs them, then turn them off again.

For a deeper treatment, see how to harden WordPress using AI and our guide to making a WordPress site AI-agent ready.

What agencies are building with WordPress MCP

Agencies lose an enormous amount of time to work that is necessary, repetitive and completely unbillable. Approving updates, publishing client drafts, clearing spam, running checks, chasing dashboards. If you look after 20 or more sites, your team repeats the same handful of tasks every week.

MCP moves that work into a conversation. Four patterns come up repeatedly once a team has MCP on more than a handful of sites.

1. An AI admin assistant on retainer

Offer clients a WordPress assistant that handles blog publishing, SEO checks, plugin audits, spam moderation and basic reporting. Much like pairing a team with a virtual administrative assistant, the AI covers the routine layer while your people handle judgement calls. Pair it with uptime and performance monitoring so the assistant has real data to report on.

2. Client reporting that writes itself

Reporting is the classic case. The AI reads the diagnostics and activity, then drafts the summary. Our guide on how to generate WordPress maintenance reports for clients covers what to include, and MCP is what removes the copy and paste from it. This slots directly into ongoing WordPress site management work.

3. AI-assisted WaaS and productised hosting

If you resell hosting, MCP becomes a product tier rather than an internal tool. Build templates with the workflows preloaded, then offer plans differentiated by how much AI access the client gets, from content-only up to full site control. Our Website as a Service platform handles the branding, plans and customer portal underneath it.

4. Productised workflows and consulting

Every agency has repeatable workflows worth packaging: prompt kits, site templates with MCP preconfigured, done-for-you assistant bundles. Beyond that there is straightforward consulting work, because most clients have no idea what is now possible. Workflow audits, MCP onboarding and custom tool design are all billable, and they position you as an integration partner rather than a supplier.

Common WordPress MCP mistakes

Most of these go wrong before a single tool ever runs, at the point where you decide which server to connect to and how much scope to hand it.

When to use WordPress MCP, and when not to

MCP is genuinely useful, and it is not the answer to everything.

The honest summary: MCP is excellent at the work you already know how to do and are tired of doing. It is poor at deciding what should be done. Keep a human on the decision and let the AI take the execution.

Which WordPress MCP route should you choose?

WordPress MCP stopped being experimental in 2026. The Abilities API is in core, the official adapter exists, and hosted options remove the setup entirely. What is left to decide is which layer you connect at and how tightly you scope it, which is a much more boring question than whether any of this works.

If you are a developer on 6.9 or later, the MCP Adapter is the core-native path and worth learning. If you run an agency and your problem is forty client sites rather than one, start at the account layer and work down. Either way, connect at read scope, prove it on staging, and widen deliberately.

Frequently asked questions