Last week I participated at the CloudFest Hackathon, where I was leading a project to bring the Model Context Protocol (MCP) to WordPress. I was intrigued by the challenge to try to bring this “USB-C port for AI applications” to the space I know best.
The goal of this project was to enable AI-powered WordPress development by implementing MCP in the WordPress ecosystem and exposing it through WP-CLI, which is the official command line interface for WordPress. So for once I was working on server-side AI and not Web AI.
WordPress development workflows currently lack seamless integration with AI capabilities, particularly during local development. While REST API endpoints enable AI interactions with live sites, developers working with local WordPress installations have limited options for AI-assisted content creation and site management.
This project transforms WordPress into an MCP Server and WP-CLI into an MCP Host through a new package, enabling direct AI interactions with WordPress installations during development. This approach provides developers with powerful AI capabilities without requiring a live site or REST API endpoints.
I also tried explaining it in simpler terms in my pre-event interview with Christian Taylor:
MCP in action
Despite MCP being bleeding edge technology, the team managed to implement a very powerful server implementation that exposes any REST API route and WP-CLI command as function calling capabilities. A separate MCP client then is able to connect to this and other servers. I say bleeding edge because MCP is brand new and lacks thorough documentation and tooling in anything that’s not Python or JavaScript. The PHP MCP SDK we used is just an AI-generated clone of the Python version and could definitely be improved. Also, the protocol itself is still undergoing big changes, as indicated by this highly relevant RFC, which was published right after the Hackathon concluded.
In addition to the MCP part, we implemented and further refined an LLM-agnostic CLI command for natural language processing. The project team used the AI Services plugin for this, but LLPhant would be a similar alternative. And ideally this would support local models too.
In short, we set out to enable use cases such as $ wp ai "Summarize my latest GitHub issues in a new blog post", where the GitHub tool could be offered by one MCP server, and the blog post tool by another.
The team even implemented image editing capabilities using the brand new Gemini 2.0 Flash native image generation that was just released days before the event. Here you can see it in action:
Hackathon experience
It was my fourth time at the CloudFest Hackathon and the second time leading a project, but the first time doing it myself. Before the event I set up a basic GitHub repository as a starting point, but didn’t have much time to do prepare more (e.g. by creating some issues or diagrams). Then, after the project patch at the start of the hackathon, I was overwhelmed by the interest in this project. We ended up being 10 people from diverse backgrounds, which was really nice!
The hackathon team:
- Myself
- Jan-Willem Oostendorp
- Joost de Valk
- Marco Chiesi
- Matt Biscay
- Moritz Bappert
- James Hunt
- Tome Pajkovski
- David Mosterd
- Milana Cap
Thanks to these wonderful people we won the social media master award thanks to our outreach during the event (special thanks to Milana for a ton of great tweets!), and overall won second place!
Conclusion
I am really excited about the future potential of the project and keen to continue working on it in some form or another. Especially the thought of implementing the MCP server in WordPress core or in a plugin is very promising.
People interested in the project are encouraged to check out the GitHub repository to collaborate.
Leave a Reply