Skip to content

Build1 publisher3 min readPublished

Adding up WordPress.com's documented MCP steps for a five-page site reaches 22 calls

A Thai-language walkthrough estimates 15 to 20 MCP calls to build a ceramics studio site on WordPress.com. Counting each documented operation, including one status update per item to publish, gets to 22.

The Engineer · Build desk

Illustration accompanying Adding up WordPress.com's documented MCP steps for a five-page site reaches 22 calls

What happened

  • Nokka published the second part of an MCP series on dev.to on 19 September 2026, running one brief, a five-page site for a ceramics studio, through the MCP servers of four content platforms.
  • WordPress.com routes the whole job through one tool, wpcom-mcp-content-authoring, whose three actions are list to see what is available, describe to see the data shape, and execute to run the operation.
  • The call counts in the piece were read off each vendor's documentation rather than measured from a run, and an action that bundles several steps into one call was counted once.

Compiled by The EngineerSomething wrong?How this is made

Why it matters

  • exposure A create that returns successfully can still have had blocks or HTML stripped, so an agent that ignores _content_warnings will leave pages missing sections that nobody flagged to the site owner.
  • decision Anyone wiring an agent editor has to choose between pages.update, which overwrites the whole page, and the page-sections calls that target one block, and that choice decides whether an edit destroys copy the agent never read.
  • cost Budgeting per-call spend or rate limits off the published 15-to-20 band will run short, because the band depends on bundling and the individually counted operations come to 22.
  • constraint Agent code written against WordPress.com's list/describe/execute triple does not carry over to Storyblok's search/describe/door cycle, so a shared protocol still leaves the backends non-interchangeable.

Call execute on a WordPress.com write without the `user_confirmed` parameter and nothing happens [14]. The flag is mandatory on writes and deletes alike, and permanently removing a category or a tag needs a further confirmation on top [14]. Deletes go further still. Across posts, pages, comments, media, categories and tags, the server makes the agent fetch the target first and display its name, its author and the number of entries it affects [16].

A create can return successfully and still have lost content. WordPress may drop blocks or HTML it does not support, without asking. The notice arrives as a `_content_warnings` field on the result, and Nokka's walkthrough says to read it after each create and each edit [15].

Nokka puts the whole site at roughly 15 to 20 calls, with a single page at 2 to 3 including pattern lookups [17]. Take the low end and five pages is 10. Two context reads before writing anything, `theme.presets` and `blocks.allowed`, make 12 [8]. `categories.create` and `tags.create` take it to 14, `posts.create` to 15, `media.create` to 16 [11][12]. Publishing is a status update per item, so five pages and one post add six, and the total lands at 22 [13][22]. That is two past the top of the published band [23]. The method note is what reconciles them: an action that bundles several steps into one call is counted once [6].

The 15 figure transfers to your own run only if `patterns.list` and `patterns.get` supply the layout so a page is effectively one create [10]. It also needs publishing to bundle instead of walking item by item, and a stripped block never sending the agent back for a second edit.

The byline discloses that the article was drafted by the glm-5.3 model through Nous Research's Hermes Agent and checked by Nokka [2]. Most comparison tables carry less provenance than that.

Storyblok's manifest starts from a different premise. It does not expose the 160-plus endpoints of the Management API as one tool each. It ships seven tools and a three-step cycle: `search` to find an operation, `describe` to get its parameters and data shapes, then execution through one of three doors [18]. The doors are separated by risk: `execute_readonly` for reads, `execute_mutating` for creates and edits, `execute_destructive` for deletes [19]. Nokka ties that separation to MCP security guidance on keeping write and delete levels apart [20].

The supplied text breaks off inside the Storyblok section, before the condition attached to `execute_destructive`, and the Sanity and Webflow walkthroughs are not in it [21]. Both servers make the agent look an operation up before it can run one, and the lookups differ in name, in gate count and in confirmation semantics. A harness written against WordPress.com's list/describe/execute triple does not point at Storyblok unchanged [7][18].

What to watch

  • Whether the later instalments show Sanity and Webflow using a third and fourth manifest shape, or converging on the same search/describe/execute cycle.
  • Whether anyone runs the five-page brief for real and reports measured call counts against the 15-to-20 documentation estimate.
  • Whether the AI image generation the docs place inside WordPress.com's content-authoring tool stays bundled with the write path.
Loading claim ledger
Loading source directory links
Loading share composer
Loading topic controls
Loading related stories