WordPress Targets 7.3 for Built-In AI Agent Access

WordPress’s AI team wants you to connect ChatGPT, Claude or any agent you like to your site and let it handle the jobs you’d normally do in your dashboard. WordPress 7.3 is the target.

It follows early tests, reported in the team’s weekly summary, in which agents finished none of the tasks through the MCP Adapter.

The four-part plan

The AI team’s plan says that today’s routes, from WP-CLI to the REST API, often require local access or more advanced configurations.

Here’s the split:

• MCP: rapidly iterate on the MCP Adapter plugin alongside WordPress 7.2, then propose bringing MCP functionality into core for 7.3, either through a direct merge or a bundled approach
• abilities, the actions an agent may call, with a starter set for common site jobs
• OAuth, so you approve an agent by signing in
• the whole journey, from an agent finding your site to you revoking its access

The goal for 7.3 is that you can link your agent “without complex setup or technical knowledge”.

Will it ship on time? The team calls 7.3 a target, and no release commitment was made. Core inclusion depends on technical readiness, security and the normal core review processes.

Then there’s the green light. The team’s weekly summary says Matt approved making MCP a de facto part of the WordPress experience.

The summary adds that this does not necessarily mean moving the entire MCP Adapter into WordPress core.

How it lands is still open. Bundling the plugin, using core interfaces you already have and building parts into core were all on the table, and no final implementation decision was reached.

What the benchmark found

So how well does the current connector work for you?

Badly, for now.

WordPress Agent Benchmarks, a project whose findings Andy Peatling shared with WordPress’s AI team, gave two agents everyday requests a site owner might make.

Each agent tried every request with and without the MCP Adapter, a plugin that turns WordPress abilities into tools an agent can call.

Here’s how the two setups ended:

• through the MCP Adapter: none finished
• with the owner’s login and WordPress’s built-in API: higher success rates, per the AI team’s summary

But the agents weren’t to blame. WordPress’s own abilities only read site, user and server details, and the content-create, content-update and content-delete abilities are still to be ported to the adapter.

Peatling opened an issue in the MCP Adapter repository on October 20 to close that gap, with milestones at 30%, 45%, 90% and 100% of owner requests as more abilities land.

Why it matters

You don’t have to wait for 7.3 to try agents on your WordPress site.

Still, today’s easy route means handing an agent an Application Password.

That’s what the OAuth work is meant to fix.

Would you give a contractor your admin login, or a key that only opens one door?

Either way, I wouldn’t put the MCP Adapter on a client site yet. It’s still a 0.x plugin, and right now it can’t do any of the 20 owner requests in the benchmark.

Build plugins? Then register your abilities early. In the benchmark, agents already used plugins’ own API routes, like WooCommerce’s, without being told how.

Common Questions

Will WordPress 7.3 include MCP support for AI agents in core?

Possibly. The AI team proposes bringing MCP functionality into core for WordPress 7.3, either through a direct merge or a bundled approach. Inclusion still depends on technical readiness, security and the normal core review processes.

Why did AI agents fail WordPress tasks through the MCP Adapter?

Mostly because the MCP Adapter does not bundle the abilities needed for common WordPress operations, according to the AI team’s weekly summary. WordPress’s own abilities only read site, user and server details, per WordPress Agent Benchmarks.

How do AI agents connect to a WordPress site today?

Agents can use WP-CLI, the REST API and file management, though these often require local access or more advanced configurations. Through the MCP Adapter, you manually manage authentication with Application Passwords. The AI team wants OAuth instead, so you approve an agent by signing in.