WordPresscom lets AI agents write publish and manage your site


For most of the past six months, connecting an AI agent to your WordPress.com site has meant giving it a window. You could ask Claude or ChatGPT questions about your content, pull up site analytics, or check which posts hadn't been updated in a year. Useful, but fundamentally passive.

On Friday, Automattic added a door.

WordPress.com has launched write capabilities for its Model Context Protocol (MCP) integration, enabling AI agents to create and modify content directly on your site.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!

The update adds 19 new operations across six content types: posts, pages, comments, categories, tags, and media. From a single natural language prompt, an agent can draft and publish a post, build a landing page using your theme's block patterns, approve and reply to comments, reorganise category structures, or fix missing alt text across your entire media library.

The underlying architecture, MCP, an open protocol that standardises how applications provide context to large language models, was first introduced on WordPress.com in October 2025. At that point it was read-only: agents could query your site but could not touch it.

A second update in January 2026 added OAuth 2.1 authentication, making it simpler to connect AI clients securely. In February, Automattic launched an official Claude Connector, again read-only at the time. Today's write capabilities are the step the platform has been building towards.

The feature is designed around explicit human approval. Before creating, updating, or deleting anything, the agent describes exactly what it plans to do and asks for confirmation. New posts default to draft status, giving users a chance to review before anything goes live; modifying a published post triggers a warning that changes will be immediately visible.

Deletions of posts, pages, comments, and media send items to the trash, where they are recoverable for 30 days. Categories and tags, which WordPress cannot trash, trigger an additional confirmation warning that deletion is permanent. Every action is logged in the site's Activity Log.

User role permissions are fully enforced: an Editor can create and edit posts but cannot change site settings; a Contributor can draft but not publish.

One of the more technically interesting aspects of the implementation is theme awareness. Before creating a page or post, the agent can read the site's design system, colours, fonts, spacing, block patterns, and generate content that inherits those specifications.

The write capabilities are available today on all WordPress.com paid plans. Users enable them through the MCP dashboard at wordpress.com/me/mcp, toggling on the specific operations they want to permit on each site.

Compatible clients include Claude, ChatGPT, Cursor, and any other MCP-enabled tool. WordPress.com powers a significant share of the web, according to figures presented at Automattic's State of the Word event in December 2025, WordPress runs more than 43% of all websites globally and holds a 60.5% share of the content management system market.

The scale at which write-capable AI agents can now operate across that infrastructure is considerable.

The MCP ecosystem has been expanding rapidly. The WordPress MCP Adapter, which enables similar functionality on self-hosted WordPress installations, has been moving toward inclusion in WordPress Core.

Automattic's other products, including WooCommerce and Beeper, have their own MCP implementations. The pattern, standardised AI agent access to application functionality, rather than one-off integrations, is becoming an architectural assumption rather than an experiment.

For WordPress.com users, the practical question is trust. Giving an AI agent write access to a production site is a different proposition from asking it to summarise your traffic. Automattic has leaned into this explicitly, making the approval model the centrepiece of the announcement and granular per-operation toggles the default configuration. 

Also tagged with