WordPress AI plugin 1.5.0 and MCP Access

· by Andrei

You log into WordPress, open a screen under Tools, and see a list of site actions you can expose to an AI agent. Not your whole admin area. Not your password. Specific abilities, one by one. That’s the interesting part of WordPress AI plugin 1.5.0.

If you’ve been hearing “AI agents can manage your site now” and rolling your eyes a bit, fair. My take: the official WordPress AI plugin is handling this in a much saner way than random tools that want blanket access and hope nothing breaks.

Still, this is an Experiment. I wouldn’t throw it onto a live store and let it roam. Start read-only, on staging first.


What MCP Access actually is

Plain English first.

MCP means Model Context Protocol. It’s a standard way for AI apps and agents to discover and use tools exposed by another system. Here, that system is your WordPress site. WordPress calls those tools Abilities.

In AI plugin 1.5.0, the new MCP Access experiment adds a real admin screen at Tools > MCP Access. According to the WordPress AI team’s 1.5.0 release post, enabling it installs and activates the MCP Adapter plugin when site permissions allow. Then you can review registered WordPress Abilities, expose or hide them one by one, and restore defaults if you mess something up.

That matters because the first question should be simple: what exactly can the agent touch?

Not everything. Good.

The adapter exposes an MCP endpoint on your site. Remote clients usually connect through a local proxy and authenticate with WordPress Application Passwords. Local use can work over WP-CLI too. The key detail from the MCP Adapter docs: Abilities are private by default, the HTTP transport requires a logged-in user by default, and each ability has its own permission check too.

So, two locks.

One more thing for developers. WordPress 6.9 introduced the Abilities API, and AI 1.5.0 updates three general Abilities to match versions coming in WordPress 7.2: core/content-query, core/users-query, and core/settings-get. Older aliases like core/read-content were removed, so older integrations can break until they’re updated.


What else 1.5.0 adds that matters for agents

MCP Access is the headline, but it’s not alone.

There’s also WebMCP. This lets compatible AI-powered browsers work inside the block editor through an experimental interface. They can read the post, update the title, insert or move blocks, change block attributes, save, and publish. You see edits live in the editor.

Version 1.5.0 also makes settings clearer. Screens now show which experiments run automatically, and an info icon explains when each one runs, what it sends to the AI provider, and how many requests it makes. That’s a good change because hidden background requests are how costs get wierd fast.

And if you already use the plugin, update it. This release fixes two reported security advisories and adds protections around stored connector credentials.


What a business site can realistically do with WordPress MCP

This is where people either get practical or get carried away.

Right now, the best uses are assistant-style tasks with a human still in the loop:

  • Have an AI assistant query your content and flag outdated pages, weak meta descriptions, or posts with no internal links. Then a person reviews the list and makes the edits.
  • Use WebMCP in the block editor to restructure a long article, tidy headings, or draft a page while you watch what changes.
  • Let an agent pull context from the site before helping with work in another system, like a CRM or helpdesk. That’s the kind of thing we build in AI integrations and automation projects.

For e-commerce, I’d keep expectations modest for now. The direction is obvious, routine catalog updates and content workflows once broader write abilities arrive. But in this release, the safe business value is mostly reading, suggesting, drafting, and helping a human move faster.

That’s still useful. A marketing manager with years of blog posts doesn’t need magic. They need a faster way to find stale content and weak structure without opening every page by hand.

If your goal is visibility in AI-driven search, this connects with the same bigger shift I wrote about in our AI Overviews article.


Safety setup I’d use first

If you want to test AI agents on WordPress without doing something reckless, use this setup.

  1. Update to AI plugin 1.5.0 first. If you already run it, do the update before experimenting because of the security fixes.
  2. Try it on staging. The plugin team itself recommends staging for experiments. On a live site, take a backup first.
  3. Create a separate WordPress user for the agent. Lowest role that can do the job, its own Application Password, revoke it when done. Never hand an agent an admin login.
  4. In Tools > MCP Access, expose only the Abilities you need. Start with read-only ones like content-query. I’d skip settings and users at first.
  5. Check the info icon for each experiment before turning it on. Some run automatically, so look at what gets sent to the AI provider and how many requests it makes.
  6. Turn on AI Request Logging and actually look at it. Also watch provider costs, because the plugin is free but model usage is not.
  7. For WebMCP, keep a human reviewing changes before publish. Revisions are there for a reason.

If you want help setting this up safely or figuring out whether your use case makes sense yet, contact me. And if you already have AI features live and need somebody to watch quality, connectors, and spend, that’s the kind of work we do in ongoing AI development and support.


Where this is heading, and when to get help

The WordPress AI team has already signaled what may come next in 1.6.0 and later: AI Workspace, Site Agent, Agent Users, role and user access controls, and more create, update, and delete abilities for content, media, users and terms. Important word here: exploratory. Not shipped yet.

So where does that leave you today?

If you’re a business owner, I’d test WordPress MCP now only if you have a real use case. Content audit. Draft assistance. Editor-side help. Something concrete.

If you’re running a high-stakes WooCommerce setup, I’d be stricter. Read-only first. Then decide.


FAQ

Is MCP Access safe on a live site?

Safer than giving an agent full admin access, yes. But I’d still start on staging, keep it read-only at first, and use a separate low-privilege user with its own Application Password.

Do I need to pay for it?

The WordPress AI plugin itself is free. You still need an AI connector and your own provider account, and the provider bills for usage.

What can an AI agent do on my WordPress site right now?

Right now, the sensible use is mostly reading and assisting. With MCP Access it can use the Abilities you expose, and with WebMCP a compatible AI-powered browser can edit inside the block editor while you review.

Does it work with the Classic Editor?

No. The plugin requires the Block Editor, and the plugin page says Classic Editor is not supported.

← Back to blog