Gutenberg MCP: how an AI writes real block content (not HTML soup)

Gutenberg MCP means giving an AI agent tools that speak the block editor’s native language: registered block types, theme patterns, and structured block markup — instead of letting it paste a wall of HTML into a post. The difference shows up the moment a human opens the page in wp-admin. Block-native content is editable; HTML soup is a “Classic” blob you have to untangle. ReDock’s MCP tools are built block-native from the ground up.

The HTML soup problem

Ask a raw language model to “write a landing page” and it will produce HTML. Often decent HTML. But WordPress doesn’t store pages as HTML — it stores them as blocks: serialized structures with comments like <!-- wp:heading --> carrying attributes the editor understands.

Paste model-generated HTML into a post and one of two things happens. Either it lands as a single Classic block — uneditable as blocks, unstyled by your theme’s block settings — or an importer makes a rough guess at block boundaries. Both paths leave the human who inherits the page worse off. The content exists, but the site’s editing model has been bypassed.

The fix isn’t a smarter model. It’s a better tool surface.

What block-native tools look like

ReDock exposes the whole app as MCP tools on localhost:47200, and the WordPress authoring tools are deliberately structural:

• list_block_types / get_block_type — the agent discovers which blocks the site actually registers, with their attributes, before writing anything.
• list_patterns / insert_pattern — the theme’s real block patterns, insertable as units.
• get_post_blocks / patch_post — read a post as a block tree, modify specific blocks, write it back. Surgical edits, not full rewrites.
• get_post_rendered — read back the final rendered output to verify the change actually took.

An agent working through these tools can’t produce soup, because the tools don’t accept soup. It composes from the same vocabulary a human editor uses.

Patterns first, blocks second

The strongest results come from a patterns-first approach. Themes ship curated block patterns — a hero, a pricing table, a testimonial row — that already carry the theme’s spacing, colors, and typography. ReDock’s authoring guidance steers agents to reach for those first, then adjust content inside them, rather than assembling pages block by block from scratch.

The practical effect: pages an agent builds look like they belong to the site, because structurally they do. They also stay fully editable — a human can open the page and tweak a heading without archaeology. We go deeper on this workflow in AI edits your WordPress site: patterns-first authoring explained.

Verify, then report

Block-native writing enables something HTML pasting never could: verification. After a patch_post, the agent reads the block tree back and confirms the edit survived. For visual changes it can request a screenshot of the rendered page. ReDock’s tools return a verified block with every result, and mutations carry an undo hint — a WordPress revision or a database pre-image — plus an audit ID.

That loop matters because models make mistakes. A tool surface that makes mistakes detectable and reversible is what turns “AI wrote something” into “AI made a change I can trust or roll back.”

All of this runs against real local WordPress sites — each at a real https://name.test domain with a trusted certificate, provisioned in one click with pinned PHP and its own MySQL database. No Docker, no terminal. The first-site guide shows the whole flow, and Free local WordPress environment with real HTTPS explains why local is the right place for agent experiments.

Why not page builders?

Block-native tooling is also why ReDock is Gutenberg-only. Page builders like Elementor store content in their own formats, without an equivalent open, structural tool surface for agents to drive. Gutenberg’s serialized block grammar is documented, inspectable, and diffable — which is exactly what an agent needs. We lay out the full comparison in Elementor vs Gutenberg in the AI era. To be clear: ReDock has no Elementor integration.

FAQ

Will pages written by the agent be editable in the normal block editor?

Yes — that’s the point. The agent writes the same serialized blocks the editor produces, using the site’s registered block types and patterns. Open the page in wp-admin and every block is individually selectable and editable.

Which MCP clients can use these tools?

Any MCP client: Claude Desktop, Claude Code, Cursor, or anything else that speaks the protocol, pointed at localhost:47200. Configuration for each client is in the MCP docs. ReDock also ships a built-in agent and a local-model catalog if you’d rather not use a cloud model at all.

Does this work with any theme?

The tools read whatever the active theme registers — its block types, patterns, and design tokens. Themes with richer pattern libraries give the agent more to compose with, but the mechanism works with any block theme.

If you’d rather your AI wrote blocks than soup, download ReDock — ReDock is free, and your first block-native site is one click away.