I've been building something called Stonewright because I kept running into the same problem when using Codex, Claude Code, Cursor, and other AI coding agents on actual WordPress projects.
Giving an agent access to WordPress is easy.
Being reasonably confident about what it changed afterwards is the harder part.
So instead of making another thin REST wrapper, I started building the workflow around the write itself:
inspect β plan/dry-run β approve when needed β snapshot β write β read back β verify β audit/restore
Stonewright now works across Elementor, Gutenberg/FSE, WooCommerce, ACF, media, content, SEO integrations and WP-CLI.
Elementor is still probably the deepest part of the project because that's where I personally needed it most.
For example, after an Elementor mutation it can snapshot the document, perform typed writes against the actual Elementor schema, regenerate CSS for the specific post/loop target, read the result back and then require explicit frontend verification instead of treating a successful API response as proof that the page is fine.
Gutenberg also has a browser-assisted finalization path now for blocks that need the actual editor runtime to serialize correctly.
There are currently 389 Plugin abilities and 101 Direct tools, but they're not all dumped into the model context. Stonewright exposes smaller task-aware surfaces depending on what the agent is doing.
It's still a public beta and I absolutely don't claim automation can't break things. The goal is to make failures easier to catch, understand and recover from.
Free and open source:
https://github.com/cosmincraciun97/stonewright-wp-mcp
If you find something stupid, please open an issue. That's genuinely more useful to me right now than a star.