Ostheimer Cockpit for WebMCP

By Andreas Ostheimer

0+ active installationsVersion 1.1.0

Let AI agents draft and manage WordPress posts via WebMCP or MCP — with a human review step before anything is published.

Description

Ostheimer Cockpit for WebMCP turns your WordPress site into a set of structured tools for
AI agents, with a human review step before anything is published. There are two ways to use it:

  • In the browser (WebMCP, experimental): 16 tools are registered in your
    browser tab through the experimental
    WebMCP Draft Community Group Report,
    which is not a W3C Standard. It needs a browser or extension that implements
    WebMCP (Chrome 149+ with an origin trial or a local testing flag). This way
    needs no separate MCP server and no plugin API key.
  • With an MCP client (optional, off by default): four of those tools can
    also be offered to MCP clients such as Claude or ChatGPT through the official
    MCP Adapter plugin, without a browser. See “Also for MCP clients” below.

The agent drafts; the human approves.

How it works

  1. Install and activate the plugin, then open the normal WordPress dashboard
    or the Cockpit for WebMCP page — both register the agent tools in your
    browser. A native dashboard box shows the live status and quick links.
  2. Enable experimental WebMCP support. Chrome 149+ offers an origin trial and
    a local testing flag; the cockpit links to the current Chrome instructions.
  3. Tell your browser’s AI agent what to do, for example:
    “Write a draft post about pumpkin recipes and submit it for review.”
  4. The agent drafts the content and submits it with the status
    Pending Review — it never publishes on its own.
  5. You review the draft in the review queue and approve it with one click.

Tool registration and WordPress requests run in your browser tab and your own
WordPress installation. The plugin itself does not send content to the plugin
developer or another third-party service. Your chosen browser, AI agent, or
extension may process prompts, page content, tool inputs, or tool results on
external systems under its provider’s terms and privacy policy.

Also for MCP clients (optional, off by default)

Besides WebMCP in the browser, four tools can be offered to server-side MCP
clients such as Claude or ChatGPT: list_posts, get_post, create_post and
update_post. They are registered as WordPress Abilities (WordPress 6.9+) and
reach the client through the official MCP Adapter plugin, no browser needed.

  • Opt-in: off until an administrator ticks the box under Tools → Settings
    (see Installation). Without the MCP Adapter plugin nothing becomes reachable.
  • Same review boundary: an MCP client can only create drafts and content in
    the pending review status and update drafts and content in that status.
    Published or scheduled content stays untouched; publishing remains a human
    action in the review queue.
  • Same protections: the capabilities of the user the client signs in as,
    the Post types setting, the rate limit and the activity log apply. The log
    marks entries that came through an ability with [via ability].
  • Errors carry the stable code in the message, for example
    [wmcp_review_required].

Agent tools included

  • create_post / update_post — common core post fields: title, content,
    excerpt, slug, author, scheduling date (planned posts), comment and ping
    status, password protection, sticky, post format, page attributes
    (parent, template, order).
  • Featured images in one step — the agent passes an image URL or uploads
    self-generated images as Base64 (PNG, JPEG, WebP, GIF), the plugin puts
    them into the media library and sets the featured image.
  • Taxonomies by name — categories_names / tags_names resolve names to
    terms; missing tags are created automatically.
  • SEO fields are written to Yoast SEO or Rank Math automatically when
    active, so agents can set meta titles and descriptions where your SEO
    plugin actually reads them.
  • list_posts, get_post, get_post_markdown — search and read content in
    agent-friendly formats.
  • list_comments, moderate_comment — triage the comment queue: the agent
    sorts pending comments and spam, you decide what gets approved. Author
    email and IP address are never handed to the agent.
  • list_revisions, compare_revisions, restore_revision — read a post’s
    revision history and roll back to an earlier one. Restoring follows the
    same review boundary as update_post (published or scheduled content
    cannot be touched) and is itself reversible, because WordPress saves the
    pre-restore state as a new revision before restoring.
  • list_terms, get_site_info, get_activity — context and audit trail.
  • upload_media, upload_media_base64 — media library uploads.
  • Developers can register their own agent tools with the
    webmcp_cockpit_tools filter — anything a PHP callback can do becomes an
    agent tool.

Safe by design

  • Agent content lands in a dedicated review status, invisible to visitors.
  • Publishing is only possible from the review queue, and only for users with
    the publish capability.
  • Successful and failed agent tool calls plus workflow actions are written to
    a controllable activity log with timestamp and user. Read entries contain
    metadata only, never post content, search terms, tool inputs or tool results.
  • Session-bound authentication with a dedicated workflow nonce; per-user rate
    limiting on write operations.
  • Works with caching plugins (Autoptimize, WP Rocket, LiteSpeed Cache,
    W3 Total Cache, SG Optimizer) and with the Classic Editor and Gutenberg.

For developers

The cockpit is extensible: register custom tools via
apply_filters( ‘webmcp_cockpit_tools’, $tools ) with a name, description,
JSON schema and PHP callback. Custom tools appear alongside the built-in ones
in the browser agent’s tool list.

Installation

Install and activate the plugin first, then choose a way (or both).

  1. Upload the ostheimer-webmcp-cockpit folder to /wp-content/plugins/, or install
    it through the WordPress plugin directory.
  2. Activate the plugin through the Plugins screen.

In the browser (WebMCP)

  1. Open the normal WordPress dashboard or Tools → Ostheimer Cockpit for WebMCP.
    The dashboard box shows the live registration status and quick links.
  2. Keep that dashboard or cockpit tab open and talk to your browser’s AI
    agent.

Requirements: WordPress 6.0+, PHP 7.4+, and a browser or extension that
implements the experimental WebMCP API (document.modelContext). Chrome 149+
offers WebMCP through an origin trial or the
chrome://flags/#enable-webmcp-testing flag for local testing. WebMCP is in
early preview and its API may change.

With an MCP client (optional)

This offers four tools (list_posts, get_post, create_post, update_post)
to MCP clients. It needs WordPress 6.9+ and is off by default.

  1. Install and activate the
    MCP Adapter plugin (checked
    against version 0.7.1).
  2. Go to Tools → Settings (administrators only), tick Offer list_posts,
    get_post, create_post and update_post as WordPress Abilities (experimental)

    and save. If the box is greyed out, your WordPress has no Abilities API.
  3. Use a WordPress user with only the capabilities the tools need
    (edit_posts, for example an Author), not an administrator. The client
    acts as this user.
  4. For a remote client, create an application password for that user under
    Users → Profile → Application Passwords. WordPress offers application
    passwords on HTTPS sites (and in local environments).
  5. The endpoint of the plugin’s MCP server (server ID webmcp-cockpit) is
    https://example.com/wp-json/webmcp-cockpit/mcp. If your site uses plain
    permalinks, use https://example.com/?rest_route=/webmcp-cockpit/mcp.
    Replace example.com with your domain. The tools appear to the client as
    webmcp-cockpit-list-posts, webmcp-cockpit-get-post,
    webmcp-cockpit-create-post and webmcp-cockpit-update-post.
  6. Enter the endpoint and the user’s credentials in your MCP client, as
    described in that client’s own documentation. For clients that start a
    local command instead, the MCP Adapter ships a STDIO bridge; run it on the
    server with wp mcp-adapter serve --server=webmcp-cockpit --user=USERNAME.
    The MCP Adapter’s
    CLI Usage Guide
    shows example client configurations for both ways.

The MCP path is new and has been tested at the protocol level (STDIO and HTTP
requests) but not yet with a specific client application such as Claude or
ChatGPT. Only four of the 16 tools are available this way; the others stay
browser-only.

FAQ

Can the AI agent publish posts on its own?

No. By design the agent can only create or update drafts and content in the
dedicated review status. Published or scheduled content cannot be changed or
moved by an agent. Publishing is a separate, capability-protected action that
a human performs in the review queue.

Which AI agents work with this?

In the browser: any compatible browser agent or extension that uses the
experimental WebMCP API (document.modelContext). The Model Context Tool
Inspector can list and manually call tools for testing. Its natural-language
prompts may be processed by an external AI provider, so review the extension’s
current terms before using private content.

Through MCP (optional): MCP clients that can connect to a WordPress MCP
endpoint, for example Claude or ChatGPT, via the MCP Adapter plugin, with four
of the tools. See Installation for the steps. This path has not yet been tested
with a specific client application.

Can the agent moderate comments?

It can triage them. list_comments shows pending, approved or spam comments,
and moderate_comment approves, unapproves or marks a single one as spam.
Both require the moderate_comments capability, so a subscriber or author
account cannot touch the queue.

Comment listings never include the author’s email address or IP address. The
agent does not need either to sort spam from real comments, and both would
leave your site with the response.

Can the agent restore an older revision?

Yes, but the same review boundary applies as everywhere else: list_revisions
and compare_revisions need the edit_post capability on the post the
revision belongs to, and restore_revision can only act where update_post
could also write — drafts and content in the pending review status. A
published or scheduled post’s revisions can be listed and compared, but not
restored by an agent. Restoring is itself reversible: WordPress saves the
pre-restore state as a new revision first, exactly as it does when you
restore a revision by hand in wp-admin.

Is it safe to let an MCP client write to my site?

The same limits as in the browser apply, and the option is off by default. An
MCP client acts as the WordPress user whose credentials it uses (for example an
application password), so give it a user with only the capabilities it needs.
It can create drafts and content awaiting review and edit drafts and content
awaiting review; it cannot publish, schedule or change published content, and
cannot delete anything. Only the four tools listed above are offered, without
media uploads or URL downloads. Every call counts against the per-user rate
limit and appears in the activity log with the marker [via ability]. You can
revoke access at any time by deleting the application password or switching
the option off. Requires WordPress 6.9+ and the MCP Adapter plugin; on older
versions the option is shown as unavailable and the browser tools keep working.

Does it send my content to external services?

The plugin itself does not send your content to the plugin developer or another
third-party service. Tool calls reach your own WordPress REST API using your
logged-in session. However, the browser, AI agent, or extension you choose may
send prompts, page content, tool inputs, or tool results to its provider.
Review that provider’s terms and privacy policy before using private content.

Do I need an MCP server or API keys?

In the browser, no separate MCP server or plugin API key is required. The tools
live on the WordPress admin page and the browser mediates between the agent and
that page. The browser agent itself may still require its own account or
configuration.

For the optional MCP client path you need the MCP Adapter plugin (it provides
the server endpoint) and the credentials of a WordPress user, for example an
application password. The plugin itself still needs no API key.

Which browsers are supported?

Chrome 149+ exposes experimental WebMCP through an origin trial or the
chrome://flags/#enable-webmcp-testing flag for local testing. Other browsers
or extensions can work if they implement the current draft. WebMCP is an early
preview, not a W3C Standard, and support may change. WordPress continues to
work normally when WebMCP is unavailable.

Does it work with Yoast SEO / Rank Math?

Yes. SEO title, meta description and canonical URL written by the agent go
directly into the active SEO plugin’s fields. Without an SEO plugin, portable
custom fields are used.

What is the upload size limit for agent images?

Base64 image uploads accept PNG, JPEG, WebP and GIF up to
min( wp_max_upload_size(), 10 MB ) of decoded data. The JSON request body is
roughly 1.37× the image size, so your host’s post_max_size must be at least
that; the error message names the exact limit when it is exceeded.

Does it work with WooCommerce or custom post types?

Posts and pages are supported out of the box. On the plugin’s settings page
(Tools → Settings) you choose which of the two agents may work with;
content of a type you switch off stays out of reach, drafts, revisions and
comments included. Additional post types and capabilities can be added with the
webmcp_cockpit_tools filter — for example WooCommerce product tools.

Who can use the agent tools?

Only logged-in users with the edit_posts capability (authors, editors,
administrators). Media uploads additionally require upload_files,
publishing requires the publish capability. Guests and subscribers get
nothing.

Screenshots

Changelog

1.1.0

  • New, optional and off by default: four tools (list_posts, get_post,
    create_post, update_post) can also be offered as WordPress Abilities, so
    server-side MCP clients such as Claude or ChatGPT reach them through the
    MCP Adapter plugin without a browser. Enable it under Tools → Settings.
    This is a market test; the other twelve tools stay browser-only for now.
  • The abilities do not re-implement any rule. Each call is dispatched to the
    same REST routes the browser uses, so capabilities, the Post types setting,
    the review boundary (published and scheduled content stays untouched),
    the rate limit and the tool audit apply unchanged. create_post accepts
    only wmcp_review or draft; a client cannot set publish.
  • MCP error results contain the stable code in the text, for example
    [wmcp_review_required]. The MCP Adapter otherwise passes only the message.
  • The activity log marks entries written during an ability call with
    [via ability], so browser and MCP activity can be told apart. No database
    change: the marker is part of the existing message column.
  • Left out of the abilities on purpose: featured_image_url (the server would
    fetch an arbitrary URL for a remote client) and all upload tools. A
    featured_media_id of an existing image still works.
  • Works unchanged on WordPress without the Abilities API (before 6.9): nothing
    is registered and the settings page shows the option as unavailable.
  • Adds tests/abilities-cases.php (all four tools, schema hardening, review
    boundary, post types, rate limit, capabilities, nonce, audit, absence of the
    Abilities API).

1.0.14

  • Closes two gaps found in a full audit of the plugin’s own guarantees.
    Trashing a post through the REST route did not check the review
    boundary, so published or scheduled content could be moved to the trash
    by an authenticated request even though every other write path refused
    it. Trashing now follows the same rule: drafts and pending review only.
  • The status transition route was the only writing route that did not
    count against the rate limit. It does now.
  • Neither gap allowed anyone to touch content they could not already edit
    or delete with their own WordPress capabilities.
  • Adds tests/guard-cases.php, which checks every documented guarantee —
    review boundary, post types, meta allow-list, rate limit, upload limits,
    capabilities, nonce, tool audit and the custom tool route — and proves
    each check by failing when the guard is removed.

1.0.13

  • Adds revision tools: list_revisions shows a post’s saved revisions
    (ID, date, author display name, short excerpt — no email addresses),
    compare_revisions returns two of them in full (title, content, date,
    author) for the agent to compare itself, and restore_revision rolls a
    post back to one.
  • restore_revision reuses the exact review-boundary check update_post
    already enforces: it can only write where an agent could also update the
    post directly, so published or scheduled content stays untouched. The
    action is reversible — WordPress saves the pre-restore state as a new
    revision before restoring, the same as a human restoring from wp-admin.
  • All three tools are rate limited like other write/read calls and appear by
    name in the activity log’s tool audit, including failed calls.
  • Fixes the Post types setting, which was saved but never enforced:
    with only Posts selected, agents could still create pages and read, edit,
    move, trash or restore revisions of existing page drafts. Every route now
    checks the setting — creating, reading (including Markdown), updating,
    status transitions, trashing, the three revision tools, and comment
    listing and moderation (limited to comments on allowed types). Blocked
    calls return wmcp_post_type_not_allowed (HTTP 403) with the allowed
    types, but only after capability and nonce checks passed, so callers
    without access learn nothing about the configuration.
  • get_site_info and the create_post schema now report only the allowed
    types, and create_post defaults to the first of them.

1.0.12

  • Adds comment moderation: list_comments triages pending, approved or spam
    comments, moderate_comment approves, unapproves or spams a single one.
    Both require the moderate_comments capability; the moderation call is
    rate limited and recorded in the activity log like every other write.
  • Comment listings deliberately omit author email and IP address — the agent
    triages, and neither field is needed for that.
  • Both comment tools appear by name in the activity log’s tool audit,
    including failed calls such as an invalid moderation action.

1.0.11

  • Renames the plugin to Ostheimer Cockpit for WebMCP and aligns the plugin
    folder, main file and text domain with ostheimer-webmcp-cockpit.
  • Limits post-list results to posts the current user is allowed to edit, with
    regression coverage for separate author accounts.
  • Moves the plugin pages below WordPress Tools and removes the obsolete
    manual translation loader and bundled runtime translation files.
  • Makes Plugin Check warnings fail the release check.

1.0.10

  • Refines the WordPress dashboard box with a clearer action hierarchy, compact
    native buttons and consistent spacing in narrow dashboard columns.
  • Keeps the copy action beside the prompt label and gives the review queue
    primary emphasis above the two secondary links.

1.0.9

  • Registers all WebMCP tools on the normal WordPress start dashboard (/wp-admin/)
    in addition to the dedicated Cockpit page.
  • Adds a native, draggable dashboard box with live tool status, a copyable
    example prompt, review count, and links to the Cockpit and activity log.
  • Adds server-side regression coverage for the dashboard widget and asset host.

1.0.8

  • Adds a complete metadata-only audit trail for successful and failed WebMCP
    tool calls, including read tools such as list_posts and get_post.
  • Adds administrator filters for read/write and tool/workflow events, bulk
    selection, and nonce-protected deletion of selected, filtered or all log
    entries. Agents cannot delete the audit log.
  • Adds an automatic database migration for the activity type and tool-name
    columns when updating the plugin in place.

1.0.7

  • Enforces the human review boundary: browser agents cannot update, publish or
    unpublish published or scheduled content; only the review queue can publish.
  • Completes the create_post and update_post schemas for author, scheduling,
    password, sticky, post format, page attributes and canonical URL fields.
  • Keeps agent-facing source output in English and clarifies that the activity
    log records agent write actions.

1.0.6

  • Tool errors now return a machine-readable signal (isError: true plus a
    stable code, following the MCP CallToolResult convention) instead of
    relying on localized text prefixes. The readable, translatable message
    stays unchanged.
  • Clarifies that WebMCP is experimental and distinguishes the plugin’s own
    data handling from external processing by a chosen browser or AI agent.

1.0.5

  • Sharper agent-facing tool descriptions: http(s)-only URLs for
    upload_media, formats/size limit/filename handling for
    upload_media_base64, precedence and failure behaviour for
    featured_image_url/featured_media_id, and the warning behaviour of
    create_post when a featured image fails.
  • Regression suite (tests/) and documentation for server-side edge cases.

1.0.4

  • i18n: the plugin source is now fully English and translation-ready.
    A German (de_DE) translation ships with the plugin.

1.0.3

  • Fix: a failed featured-image import from featured_image_url (e.g. a URL
    that returns no image) no longer reports success and no longer touches an
    existing featured image; imported attachments are verified as real images
    from the downloaded bytes, failed imports are cleaned up.
  • Fix: error messages now name the originally submitted value for invalid
    media IDs (e.g. -5) and distinguish URL format errors from network/DNS
    errors.

1.0.2

  • Fix: invalid featured_media_id values now return a clear error instead of
    a silent success; attachments are validated for existence and image type.
  • Fix: rejected media uploads (wrong type, broken Base64, size limit) are now
    recorded in the activity log as media_rejected entries.

1.0.1

  • New: upload_media_base64 tool — agents can upload self-generated images
    (PNG, JPEG, WebP, GIF) directly into the media library without a public URL.
    Real MIME type is verified from the decoded bytes; size capped at 10 MB.
  • New: featured_media_id parameter for create_post/update_post — assign
    a featured image from an existing media ID (takes precedence over
    featured_image_url).
  • Tool errors now return readable messages to the agent.

1.0.0

  • Initial release: 10 agent tools for creating, updating and reviewing posts,
    full post field coverage, featured images from URL, taxonomy assignment by
    name, SEO plugin integration (Yoast SEO, Rank Math), review workflow with
    human approval, activity log, rate limiting and a PHP API for custom tools.