Pretty cool to see your WordPress site as a connector in the Claude mobile app.
It is a streamable-http MCP server by Agent Pilot plugin combined with OAuth Pilot for seamless auth.
Also works with the official MCP-adapter feature plugin.
With Agent Pilot you can create and manage agent skills in WordPress (just like blog posts) and publish them as standard Agent Plugins.
Now with OAuth Pilot you get one-click authentication for all the MCP clients (Claude, ChatGPT, etc).
No more application passwords!
Did you know that the creator of ManageWP Vladimir Prelovac also created Kagi search engine and the associated products? So cool!
The OpenAI batch API is pretty cool at 50% of the cost. Great for large scale tasks such as captioning 42k images with gpt-5-mini.
Here is what I learned:
1. there is an org-level cap of 5M enqueued tokens per model so you need to throttle the uploads to stay below that. Not sure why they don't take care of that internally;
2. each batch upload must be below 200MB of so I had to create 75 JSONL files (multiple JSON objects per line) for 42k images resized to 1536px on the long edge.
During the weekend I built a contextual search tool for a photo library with 42k images across 2.8k galleries. The results are amazing!
Turns out the AI models that generate embedding vectors from images directly are BAD at capturing the image contents, so I had to first generate captions and then generate the embeddings for the captions instead.
Total cost: $30 for both captions (gpt-5-mini) and embeddings (text-embedding-3-large with 3072 dimensions).
The FAIR standard (originally created for WordPress and built on AT protocol) would be perfect for distributing agent plugins since it covers everything from versioning, signing, discovery and packaging.
Now the Agent Plugin project is trying really hard to figure this out https://github.com/agentplugins/agent-plugins-spec/discussions
Responded to @kasparsd:
What do you think of these Claude system prompts? Sound a bit too verbose and as if AI was used to iterate on them which eventually flattens the intent and looses the nuance.
https://platform.claude.com/docs/en/release-notes/system-prompts
Trying out Agent Pilot WordPress plugin prototype to self-host host AI skills.
Let's you manage and publish agent skills without having to figure out Git or GitHub.
There is now a standard for bundling agent skills and MCP configurations as portable plugins (zip files) https://agent-plugins.org
This is amazing for self-hosting your AI tools and not needing to publish to individual vendor marketplaces.
What would be your preferred interface for creating agent skills in WordPress?
a) all front-matter fields in the editor area (ACF-style), or
b) all front-matter fields in the sidebar panel?
Thanks to the "Agent Skills Discovery via Well-Known URIs" RFC from Cloudflare we can have WordPress as the Skills CMS with the convenience of `npx skills add https://…`
I really hope we eventually arrive at an open package standard for bundling these custom AI plugins that can be self-hosted and installed using any AI GUI harness.
Currently all the nice AI things are mostly available to developers who can run CLI commands.
See https://github.com/cloudflare/agent-skills-discovery-rfc
So when is every new plugin an AI plugin?
1. Exponential — late 2029
2. Linear — by 2038.
Nice read from Casey: https://caseyrb.com/blog/ai-plugins-on-wordpress-org/
Agent skills and instructions need a CMS so that teams can collaborate on these critical assets without having to learn git and cli tooling.
WordPress is a CMS. Can we make it do this?
There is a proposal to add a wp_knowledge post type (and a matching taxonomy) to WP core https://make.wordpress.org/core/2026/06/22/merge-proposal-guidelines-built-on-knowledge/ but it is not landing soon. The idea is to test it as a plugin first.
Check your AI tool memories (under ~/.codex/memories and ~/.claude/projects). I find that most of them are actually polluting the context and not helping.
We often overlook how WordPress was THE vibe coding of the previous decades. Now there are just way more options available.
the no-tool and no-rag approach of claude code feels to be wasting tokens and working through things slower. what's your experience?
same model in cursor is way faster and productive.
The Jetpack monorepo is 767MB 🐡
Mostly because of the tracked build files (roughly 360MB) which are not even part of the current work tree.
Would you trust AI to run this?
Will it actually apply those single quotes when running it?
It created the ~ directory itself during previous runs so it is trying to clean up.
Responded to @kasparsd:
I first heard about this concept from Adam Jacob on the Changelog podcast covering his Swamp project https://changelog.com/friends/130