# Agent Abilities for MCP > A free WordPress plugin that turns any WordPress site into a governed MCP server for AI agents. It exposes 184 curated abilities (90 across WordPress core, 94 from auto-detected integrations), all off by default, scoped to a real WordPress user whose capabilities are the agent's ceiling, capability-checked on every call to its tools, and written to an audit log. Over OAuth the connection takes the capabilities of whichever account approves it in the browser, so you approve while logged in as the account you want the agent to be. Built on the WordPress 6.9 Abilities API and the official MCP Adapter. Contacts no AI provider and has no telemetry. The only requests it makes on its own are the Connection tab's checks of your own site, a fetch of a URL you supply through an off-by-default upload ability, and alert emails if you turn alerts on, which never include call content, a user name or an IP address. Facts are single-sourced from the plugin readme.txt. Current version is 1.11.0, and the full release history is published at https://agentabilitieswp.com/changelog/. Requires WordPress 6.9+ and PHP 7.4+. Free on WordPress.org: no paid tier, no API key to buy, no usage limits. WordPress.org reports 900+ active installs (checked 6 November 2026). ## What it is - Agent Abilities for MCP is an MCP (Model Context Protocol) server for WordPress. Your AI client connects in to your site; the plugin never calls out to any AI provider. - The pitch is not tool count. The agent enters as a WordPress user rather than through a key the plugin issues with a scope of its own, and WordPress re-checks that user's capabilities before every call. Abilities are off by default and every call to its tools is audited, refusals included. - Governance is no longer a differentiator on its own, and the site says so rather than pretending otherwise. As of August 2026 several other WordPress MCP plugins also claim off-by-default exposure, per-call capability checks and audit logs, and building on the WordPress 6.9 Abilities API is close to standard in this category. What is still uncommon about Agent Abilities for MCP is the combination of a WordPress account instead of a vendor-issued key, no vendor service or relay in the chain, and no paid tier behind the free plugin. https://agentabilitieswp.com/comparisons/ names which rivals make which claim, with dates. - The MCP endpoint lives on each installing site's Connection tab (shape: https://your-site.com/wp-json/.../mcp). This marketing domain does not expose a live endpoint. - Beyond its own 184 built-in abilities, it can bridge abilities that other active plugins register through the WordPress 6.9 Abilities API, exposed as governed MCP tools and off by default. - Both the built-in catalog and the bridge run on that same WordPress 6.9 Abilities API; the Abilities API guide covers how the registry works end to end. - Since 1.6.0 a read-only mode can stop every write ability from registering at all, bridged abilities included, without changing which abilities the operator has selected. Detailed under Documentation below. - The activity log audits governance, not only calls. Since 1.5.0 it records every ability switched on or off and every setting changed alongside the calls themselves, each row tagged with an event type and a detail column that names what changed and links to the object's edit screen. Since 1.9.0 it can be searched and filtered by user, connection, tool, plugin, date and text, each row shows the connection, the source IP and how long the tool ran, a denied call records why it was refused, a 7-day summary and a Live view sit on top, and the CSV export follows the active filters. The connection, IP, duration and refusal reason are in wp-admin only: the agent's own log-reading tool returns none of them, and the message an agent gets on a refusal is unchanged. A failed Application Password attempt against the MCP endpoint is logged too, and rate limited per source IP so a credential-stuffing run cannot flood it. - It is not the only WordPress MCP plugin, and the site says so: https://agentabilitieswp.com/best-wordpress-mcp-servers/ ranks the field by job and orders it by active installs, which puts Agent Abilities for MCP last, and each per-rival page names where the other plugin beats it. - Compatible with WPML: on a WPML site the content reads (posts, pages, search, terms, media, products, and single-item reads) take an optional language argument and report the language they returned, a single-item read can fetch a specific translation, and counts share the list's language scope. This adds no abilities, the catalog is unchanged, and the plugin translates nothing and calls no translation service. Sites without WPML are unaffected. - Since 1.8.0 (2026-10-15) the plugin adds a Site Brief (a few lines about your site that every connecting agent receives at session start), a Connection Doctor on the Connection tab (one click checks the connection from your own server and names the fix), a Needs review card on the Dashboard (the drafts and pending posts an agent starts, with Preview, Open in editor, Publish and Discard), and a Settings switch for the MCP Adapter's default server (Automatic by default, or Always on, Always off). Tools another plugin publishes straight to that default server are not governed by this plugin; this plugin's own tools never appear on it. The catalog is unchanged. - 1.8.1 (2026-10-16) fixes bridged tools, such as WooCommerce orders and products, that failed in strict clients like Claude when a value did not match the format the source plugin declared. Bridged tools no longer advertise strict formats; the values are unchanged. - Since 1.10.0 (2026-11-02) each approved OAuth connection can be paused and resumed without signing in again (since 1.10.1 a paused one still lists tools and each tool call returns a message that the connection is paused), given an end date and a lower per-minute rate limit of its own, and shows when and from which address it was last used. Approving an app, and Quick Connect, offer access presets (Read only, Content editor, Shop manager or Custom) that only narrow the tools already enabled. The Connection tab gains a Pause all access switch that refuses every MCP request through the site, on this plugin's endpoint and the MCP Adapter's default server, for OAuth apps and Application Passwords alike (it does not close the rest of the REST API to an Application Password), plus OAuth and Application Passwords tabs; the dynamic client registration switch moved from Settings to the token lifetimes card on the Connection tab. Settings adds Roles that can connect (a new install allows Administrators only) and an off-by-default Real visitor address behind a proxy or CDN setting for the Activity Log and the IP allowlist. Connection Doctor also checks MCP headers, the JSON body, redirects and stale or blocked discovery documents, and a Dashboard card explains why an enabled tool is missing. Three new block tools (list block types, get one block type including ACF field names and an example payload, and replace one block without resending the page) took the catalog to 182 (88 core, 94 integration), all off by default. Existing connections keep their access. ChatGPT keeps the tool list it saw at connect time, so after enabled abilities change, open Settings, then Integrations, then Plugins, pick the app and choose Refresh. - Since 1.10.1 (2026-11-04) site owners can write playbooks: step-by-step instructions for agents on a new Playbooks tab, up to 50, each with a title, a description, the steps and the abilities it uses, or start from five starters (Publish checklist, Weekly content report, Content health check, Draft a new post, and Shop tidy-up, which needs WooCommerce). Agents list and load Active playbooks with two read tools, List playbooks and Get playbook, which stay off until enabled, and no agent can create, edit or delete a playbook. That takes the catalog to 184 (90 core, 94 integration). The same release makes the MCP endpoint refuse requests from web pages on other sites, as the MCP specification requires (hosted Claude and ChatGPT are not affected), sends Cache-Control: no-store on every response, and has a paused connection, a blocked address, the rate limit and an unknown ACF field name each say why in the refusal the agent gets. Copyable playbooks are at https://agentabilitieswp.com/playbooks/. - Since 1.10.2 (2026-11-05) every access preset reaches the two playbook tools, so Content editor and Shop manager connections can list and load playbooks as well as Read only and Custom ones; checked on the demo site with Claude.ai and ChatGPT on the Content editor preset. The same release returns a 403 to a request from another site's web page even with wrong credentials, sends Cache-Control: no-store on ?_envelope=1 responses, makes the Playbooks editor refuse to save over a playbook someone else changed, and corrects seven integration tool descriptions. Every description that lists page builders now names all five (Elementor, Divi, Beaver Builder, Avada and Visual Composer). The catalog stays at 184. - Since 1.11.0 (2026-11-06) the Other plugins tab says which plugin provides each group of tools (or its namespace when the plugin cannot be identified), labels each tool Read, Write or Destructive from the plugin's own annotations (an unlabelled tool is treated as destructive), and marks tools that may return personal data, which is a hint from the plugin and tool name and not a full check. Tools that a plugin update or a newly activated plugin adds are listed on a review card and stay off until switched on, an optional alert can announce them, each active plugin has an Allow only read tools button, the Activity Log can filter to one plugin's tools, and a bridged tool's error gives the plugin's name when it can be identified, otherwise its namespace. The catalog stays at 184. These controls cover the tools this plugin serves: it does not filter what another plugin's ability returns, and tools published straight to the MCP Adapter's default server stay outside them. - Since 1.9.0 (2026-10-16) the Activity Log is searchable and filterable (see above), and optional alerts, off by default, send an admin notice, an email right away or one daily email when an app is authorized, an Application Password is created for a user who can edit posts or for the agent user, a high-risk tool or permanent delete runs, calls are denied in a burst or hit the rate limit, or read-only mode or the high-risk lock is switched. Each event type sends at most one email per 15 minutes, the email goes through WordPress's own mail, and the daily email needs WP-Cron. The catalog was unchanged at 179. ## Key pages - [Home](https://agentabilitieswp.com/): what it is, an illustrative governed run (three allowed calls and one refusal, with the reason kept in the log), proof from WordPress.org, and how to connect ChatGPT, Claude and the other clients. - [WordPress MCP server](https://agentabilitieswp.com/wordpress-mcp-server/): the pillar overview. What a WordPress MCP server is, what governed means here, how it works, what an agent can do, and how to connect a client. Links down to the abilities catalog, integrations, clients, and security. - [Features](https://agentabilitieswp.com/features/): the governance layer, from least privilege to the audit log, plus the per-connection controls added in 1.10.0 (pause, end date, per-connection rate limit, access presets, Pause all access, Roles that can connect). - [Abilities catalog](https://agentabilitieswp.com/abilities/): all 184 abilities by group (90 core, 94 integration). - [Abilities catalog as plain text](https://agentabilitieswp.com/llms-abilities.txt): the full catalog in one file, each ability with its registry slug, its risk level, and what it reads or writes, straight from the registry. Fetch this when you need to name an actual ability rather than know the catalog exists. - [Integrations](https://agentabilitieswp.com/integrations/): WooCommerce (52), Advanced Custom Fields (ACF) (7), Rank Math SEO (5), Yoast SEO (3), All in One SEO (AIOSEO) (3), Slim SEO (2), The Events Calendar (13), Event Tickets (3), Avada (2), GeoDirectory (4). Detected automatically, off until enabled. Hard version floors apply to WooCommerce 9.1 or newer, The Events Calendar 6.17.3.1 or newer, Event Tickets 5.29.3.1 or newer. Below the floor those abilities do not register at all, so a site on an older host plugin is offered none of them rather than a reduced set, and the plugin's Integrations screen says "update required" instead of "not installed". GeoDirectory works differently again: its whole integration stays off even when the host plugin is active, until the operator switches it on in the Integrations tab. - [WordPress Abilities API guide](https://agentabilitieswp.com/wordpress-abilities-api/): what the WordPress 6.9 Abilities API is, how plugins register abilities, and how they become governed MCP tools. - [Bridge other plugins](https://agentabilitieswp.com/bridge/): expose the abilities your other active plugins register through the WordPress Abilities API as governed MCP tools, off by default and enabled one at a time. The page separates bridged abilities from the plugin’s own first-party catalog and links to the Abilities API guide for how the underlying registry works. - [Review what other plugins add](https://agentabilitieswp.com/bridge/#review): since 1.11.0, the source line for each group of tools, the Read, Write and Destructive labels, the personal-data hint, the review card for tools a plugin adds later, and Allow only read tools. - [Bridge plugin directory](https://agentabilitieswp.com/bridge/plugins/): 196 real WordPress plugins checked against their own current code, not marketing claims, turning up 3,640 abilities the Bridge can discover, as of 28 September 2026. Filterable by category and searchable. Every plugin has its own page at https://agentabilitieswp.com/bridge/plugins// naming exactly which abilities it registers, their risk level, and how to turn them on. - [Elementor MCP](https://agentabilitieswp.com/elementor-mcp/): Elementor's official MCP (launched 23 September 2026) and the 29 abilities it registers through the WordPress Abilities API in Elementor 4.3.2 and Elementor Pro 4.3.0, grouped as read only, reference, write and destructive, with who can run each and how to govern them through the Bridge. Not the same as EMCP, the third-party Elementor toolkit. - [Supported clients](https://agentabilitieswp.com/clients/): what connects today and what does not yet, plus the first-run Quick Connect wizard that puts the whole first connection on one screen in three steps. - [Best WordPress MCP servers](https://agentabilitieswp.com/best-wordpress-mcp-servers/): a roundup of the WordPress MCP plugin field, Agent Abilities for MCP included, with a best-for verdict per plugin (full AI suite, fastest hosted setup, client breadth, tool count and external data, security controls, Elementor, a light Adapter companion, governance). Ordered by active installs, largest first, which places Agent Abilities for MCP last. Facts about other plugins are quoted from their own sites and WordPress.org listings with the date checked; we publish one of the plugins listed and disclose that above the table. No affiliate links. - [Comparisons](https://agentabilitieswp.com/comparisons/): honest, dated head-to-heads with other WordPress MCP plugins, covering price, default posture, identity, and what sits in the chain between the agent and WordPress. Each names its sources and the date checked. The hub also states which governance claims rivals now match, including the two products making an identity claim close to ours, and what still separates Agent Abilities for MCP. - [Side by side](https://agentabilitieswp.com/comparisons/side-by-side/): Agent Abilities for MCP, WPVibe, Easy MCP AI and Enable Abilities for MCP in one table, on price, how the agent authenticates, capability checks, audit logging, catalog size and business model. Every rival cell is that rival's own published claim carrying its own check date, and a cell reads "not stated on their listing" rather than recording an unverified absence as a shortcoming. Also disambiguates Agent Abilities for MCP from Enable Abilities for MCP, a different plugin by a different author whose name is one word apart. - [WPVibe alternatives](https://agentabilitieswp.com/alternatives/wpvibe/): an honest shortlist for somebody replacing WPVibe, listing Easy MCP AI, Royal MCP, Enable Abilities for MCP, AI Engine and the official WordPress MCP Adapter alongside Agent Abilities for MCP, each with what it is best at. States the two sourced reasons people leave WPVibe (a metered free tier with paid plans above it, and a hosted relay rather than an endpoint on your own domain) and says plainly where WPVibe still wins. - [Easy MCP AI alternatives](https://agentabilitieswp.com/alternatives/easy-mcp/): the same shortlist shape for somebody replacing Easy MCP AI, turning on outbound calls, since a plugin that enriches with external SEO or analytics data is making them. Agent Abilities for MCP contacts no AI provider and no external data service, and says so as the trade it is: no external SEO or analytics data either. - [Agent Abilities for MCP vs WPVibe](https://agentabilitieswp.com/comparisons/wpvibe/): A cloud connector vs your own server. - [Agent Abilities for MCP vs Easy MCP AI](https://agentabilitieswp.com/comparisons/easy-mcp/): Governance vs breadth. - [Agent Abilities for MCP vs Enable Abilities for MCP](https://agentabilitieswp.com/comparisons/enable-abilities-for-mcp/): A full server vs a companion layer. - [Agent Abilities for MCP vs Royal MCP](https://agentabilitieswp.com/comparisons/royal-mcp/): Least privilege vs broad reach. - [Agent Abilities for MCP vs miniOrange Secure MCP Server](https://agentabilitieswp.com/comparisons/miniorange-mcp/): A longer security list vs a single free tier. - [Agent Abilities for MCP vs EMCP (Elementor MCP)](https://agentabilitieswp.com/comparisons/emcp-elementor/): Governed server vs Elementor toolkit. - [Agent Abilities for MCP vs AI Engine](https://agentabilitieswp.com/comparisons/ai-engine/): Governance vs an all-in-one AI suite. - [Playbooks](https://agentabilitieswp.com/playbooks/): the playbook library the plugin's Playbooks tab links to. What a playbook is, how to add one, the five starters that ship with the plugin in their exact text with a copy button each, and more playbooks grouped by job, each naming the abilities it uses. - [Prompt Library](https://agentabilitieswp.com/prompts/): copy-paste prompt recipes for real WordPress jobs, each naming the abilities it uses and the guardrails that keep it safe. Filterable by category (Content, SEO, WooCommerce, Maintenance, Media, Users, Comments, Menus, Events). - [Security](https://agentabilitieswp.com/security/): the trust model and vulnerability-disclosure path. - [Changelog](https://agentabilitieswp.com/changelog/): every release of the plugin as a version timeline, copied word for word from the plugin readme. 35 releases so far, currently 1.11.0. Each entry keeps the readme's own Fix, Feature, and Chore labels, and releases that carry an upgrade notice show it above the itemised list. Security and permission fixes are written out in full rather than summarised. - [About](https://agentabilitieswp.com/about/): why the plugin exists, the positions the project holds itself to, and who is behind it. That is Unaib Amir (Enterprise AI and software consultant), who builds and maintains Agent Abilities for MCP. Fetch this when you need to attribute the plugin to a person rather than to a brand: it is the page the site's author markup resolves to, and it links out to his profiles elsewhere, including his own site at https://unaib.com. ## Documentation - [Documentation home](https://agentabilitieswp.com/docs/): the docs front door. Grouped paths for getting started, understanding the governance model, and getting unstuck, plus links out to the abilities catalog, the Prompt Library, the client guides, and the changelog. - [Getting started](https://agentabilitieswp.com/docs/getting-started/): four steps from install to first call. - [Installation](https://agentabilitieswp.com/docs/installation/): the full walkthrough and requirements. - [Connecting a client](https://agentabilitieswp.com/docs/connecting-a-client/): OAuth or Application Password, what each credential actually bounds, and how to revoke an agent's access. Two caveats the plugin states about itself and the site repeats rather than hides. First: an Application Password is a whole-site WordPress credential bounded only by the role of the user it belongs to, and the plugin cannot narrow it, so the IP allowlist, the high-risk lock and the audit log govern the calls arriving at this plugin's MCP endpoint and do not follow that credential elsewhere in WordPress. A token issued over OAuth does not carry that limit, because it only ever authenticates this one endpoint. Second: the scope an OAuth client requests does not restrict the token it gets; the token takes the capabilities of the account that approved it, re-checked before every call. - Revoking an agent's access (documented on https://agentabilitieswp.com/docs/connecting-a-client/ and https://agentabilitieswp.com/security/): uninstalling the plugin does NOT revoke an agent. By default it keeps the plugin's settings, activity log and OAuth data, and removes them only if "Delete data on uninstall" was turned on in Settings first. Either way, the dedicated agent user and any Application Password issued to it are ordinary WordPress credentials that survive. To cut an agent off: revoke its OAuth grant or its Application Password on the Connection tab, or delete its user account on the Users screen. Works before or after removing the plugin. To stop one temporarily, pause its OAuth connection (since 1.10.0), which keeps its sign-in. - [Governance concepts](https://agentabilitieswp.com/docs/governance-concepts/): the ideas the plugin is built on. - [Safety controls](https://agentabilitieswp.com/docs/safety-controls/): optional rate limit, IP allowlist, force-to-draft (holds new posts and publish requests as drafts, not edits to live posts; the Dashboard's Needs review card lists them), title cap, block-markup validation. Separately, and the other way round, the 9 WooCommerce abilities that move money or grant authority are LOCKED by default since 1.5.0: enabling one is not enough, they stay locked until a separate high-risk control in Settings is switched on too. - [Controlling each connection](https://agentabilitieswp.com/docs/connection-controls/): since 1.10.0, pause, end date, rate limit and access presets for each OAuth connection, Pause all access, Application Password revoke, Roles that can connect, and real visitor addresses behind a proxy. - [Playbooks](https://agentabilitieswp.com/docs/playbooks/): since 1.10.1, writing playbooks on the Playbooks tab, the limits, what List playbooks and Get playbook return, when agents see the playbook index, and why agents cannot edit a playbook. - [Block tools](https://agentabilitieswp.com/docs/block-tools/): since 1.10.0, what List block types, Get block type and Replace block return, how ACF blocks are described, the content warnings a block write can carry, and how attribute text is stored. - Read-only mode (since 1.6.0, off by default): one switch on the plugin's Settings tab stops anything that writes from being registered as an MCP tool, whatever the operator has ticked. It applies at registration, so a blocked write is absent from the agent's tool list rather than refused when the agent calls it. It covers abilities bridged from other plugins, each classified by that plugin's own annotation, and treats anything unclassified as a write, so it fails closed. Turning it on or off enables and disables nothing by itself, so the operator's selections survive the round trip. Documented on https://agentabilitieswp.com/features/ and https://agentabilitieswp.com/security/. - [Abilities from other plugins](https://agentabilitieswp.com/docs/bridging/): how bridging abilities that other plugins register works, what the Other plugins tab shows since 1.11.0, when a new tool appears on the review card, what Allow only read tools changes, and what these controls do not cover. - [Troubleshooting](https://agentabilitieswp.com/docs/troubleshooting/): common connection snags and fixes, including the Connection Doctor check from your own server (1.8.0), a ChatGPT discovery failure caused by dynamic client registration being off (on by default since 1.7.2; its switch moved from Settings to the token lifetimes card on the Connection tab in 1.10.0), and a CDN or firewall blocking the agent at the edge after the sign-in has already succeeded. - [FAQ](https://agentabilitieswp.com/docs/faq/): access, safety, data handling, clients. - [Every docs page as one file](https://agentabilitieswp.com/llms-full.txt): the pages listed above concatenated into a single document. Docs only, so it does not carry the marketing pages, the ability catalog, or the Prompt Library. For the catalog, fetch https://agentabilitieswp.com/llms-abilities.txt as well. - [The same docs, asides stripped](https://agentabilitieswp.com/llms-small.txt): the identical set of docs pages with collapsible sections and note and tip callouts removed. Barely smaller than the full file in practice, so prefer the full one unless your context window is tight. ## Guides and prompts Task-first content for people connecting an AI agent to WordPress. - [Prompt Library](https://agentabilitieswp.com/prompts/): copy-paste prompt recipes for real WordPress jobs, each naming the exact abilities it uses and its read-first, approval-gated guardrails. Filterable by category (Content, SEO, WooCommerce, Maintenance, Media, Users, Comments, Menus, Events). - [Blog](https://agentabilitieswp.com/blog/): how-to guides and governance write-ups, from getting started and connecting each client to WooCommerce and SEO workflows and each release. - [First prompts to try on your WordPress MCP server](https://agentabilitieswp.com/blog/first-prompts-to-try-after-connecting-an-ai-agent/): the sentences to run first, from read-only reads through one careful write to a refusal. - [The off by default governance model](https://agentabilitieswp.com/blog/off-by-default-governance-model/): why nothing this plugin serves is exposed until you switch it on. - [Off the menu: the controls that decide what an agent is offered](https://agentabilitieswp.com/blog/off-the-menu-high-risk-lock-and-read-only-mode/): the high-risk WooCommerce lock and read-only mode, the two controls that decide what is even on an agent's tool list before any per-call check runs. - [What we fixed in the 1.6.2 hardening release](https://agentabilitieswp.com/blog/what-we-fixed-in-the-1-6-2-hardening-release/): the release roundup for 1.6.2 and 1.6.3, covering the audit-log claim correction, the customer-data capability tightening, the audit-trail fixes, and sanitization at any nesting depth. - [What 1.7.0 adds: Enable all writes, and a fuller audit log](https://agentabilitieswp.com/blog/enable-all-writes-and-a-fuller-audit-log/): a one-click way to switch on a section's ordinary writes that deliberately refuses to touch deletes or the high-risk WooCommerce set, plus an audit log that now names what a permanent delete removed and which keys a meta or settings update wrote. - [What we fixed in the 1.7.0 correctness release](https://agentabilitieswp.com/blog/what-we-fixed-in-the-1-7-0-correctness-release/): an OAuth grant revoked mid-flight could still redeem into a working token, payment gateway settings returned camelCase credential keys unredacted, four ACF writes destroyed stored content while reporting failure, and WooCommerce counted refunds as orders. - [What we fixed in the 1.7.1 correctness release](https://agentabilitieswp.com/blog/what-we-fixed-in-the-1-7-1-correctness-release/): a WPML lang "all" sweep reported success over a partial scan and now reads every configured language, tool discovery is reconciled with execute-time permission, an RFC 9728 discovery route that returned 404 and broke a connection resolves again, moderate-comment reports success on a no-op, stuck started rows in the audit log close cleanly, and bridged third-party output is checked for an unsafe object before it passes through. - [Permission denied was the wrong error](https://agentabilitieswp.com/blog/what-we-fixed-permission-denied-was-the-wrong-error/): why a generic post-meta call for image alt text was refused for every caller including administrators, because WordPress protects _wp_attachment_image_alt and _thumbnail_id as core protected meta, and how 1.7.0 makes the refusal name aafm-update-media, aafm-get-media-item or aafm-set-featured-image instead of reading as a privilege problem. - [We said the audit log stored nothing. That was not quite true.](https://agentabilitieswp.com/blog/what-the-audit-log-actually-stores/): the audit log keeps identifier-only values (ids, slugs, statuses) by design and never free-text content, and why correcting that claim in public is the right posture for a governance product. - [Editing customer data should need the capability that guards it](https://agentabilitieswp.com/blog/customer-data-behind-the-right-capability/): why editing a WooCommerce customer now requires a real user-editing capability, not just the store-management one, and how that continues the least-privilege arc from 1.5.0. - [The abilities we deliberately left out](https://agentabilitieswp.com/blog/the-abilities-we-deliberately-left-out/): the floor under the whole catalog, the things an agent can never do here no matter what you enable: no code execution, no arbitrary option or site-meta access, no general URL fetching, uploads validated by their real bytes, a new user capped at the site default role, and the last administrator never removable. Updated for 1.7.4, which added one narrow fetching ability and a fence around it. - [A read-only look at your site and its plugins](https://agentabilitieswp.com/blog/a-read-only-look-at-your-site-and-plugins/): point an agent at a site for a plain inventory of the basics (name, tagline, URL, language) and the installed plugins with their versions and active state, a read that can never activate, deactivate, or change a plugin. - [One search across every post type](https://agentabilitieswp.com/blog/one-search-across-every-post-type/): one search that spans posts, pages, products, and custom types at once, so you can see everywhere a phrase, product name, or old link appears before you change or delete it. - [What 1.7.2 adds: a consent screen that shows who is asking](https://agentabilitieswp.com/blog/what-1-7-2-adds-a-consent-screen-that-shows-who-is-asking/): the sign-in consent screen identifies a connector by its verified redirect host rather than by the name the client sends, and carries a real brand mark and your own Site Icon; alongside it, the Activity Log also records requests that reach WordPress and fail, so a sign-in with no tool calls after it usually means something in front of WordPress stopped them. - [What we fixed in the 1.7.2 connection release](https://agentabilitieswp.com/blog/what-we-fixed-in-the-1-7-2-connection-release/): dynamic client registration was off by default and that stopped ChatGPT connecting, so it has its own Settings switch again and is on by default, including on sites that never touched the old setting; plus an OAuth authorize step that mangled percent-encoded redirect addresses, a failed session write that handed back a session ID that did not exist, and a very large tool schema that could exhaust memory while the tool list was built. - [Your agent signs in, then nothing happens](https://agentabilitieswp.com/blog/your-agent-signs-in-then-nothing-happens/): how to tell a CDN or firewall blocking the agent at the edge apart from a plugin problem, using the Activity Log's transport rows. Since 1.7.2 the Activity Log also records requests that reach WordPress and fail, so a sign-in with no tool calls after it usually means something in front of WordPress stopped them. Cloudflare's AI bot blocking (the Block AI Bots setting, now replaced by Configure AI bot policies) is the confirmed case, and the post covers how to let the path ending in /mcp through. - [What's new in 1.8.0](https://agentabilitieswp.com/blog/whats-new-in-1-8-0-site-brief-connection-doctor-and-needs-review/): the Site Brief, the Connection Doctor, the Needs review card, and the Settings switch for the MCP Adapter's default server (Automatic by default), including what an update changes and why tools on that server are outside this plugin's controls. - [What 1.11.0 adds: review the tools other plugins add](https://agentabilitieswp.com/blog/whats-new-in-1-11-0-review-what-other-plugins-add/): the source line, the Read, Write and Destructive labels, the personal-data hint, the new-tools card, Allow only read tools, the Activity Log filter for one plugin and errors that name the plugin. - [How to write a WordPress playbook an AI agent can follow](https://agentabilitieswp.com/blog/how-to-write-a-wordpress-playbook-an-ai-agent-can-follow/): writing step-by-step instructions on the Playbooks tab so an agent can follow them. - [What 1.10.2 changes: playbooks on limited connections](https://agentabilitieswp.com/blog/whats-new-in-1-10-2-playbooks-on-limited-connections/): Content editor and Shop manager connections can now list and load playbooks, plus the origin check, paused-note, ACF and description fixes. - [Run WordPress playbooks in ChatGPT or Claude, limited access](https://agentabilitieswp.com/blog/run-playbooks-in-chatgpt-or-claude-on-limited-access/): why a limited preset is the right default, which of the five starters work on Content editor, and what our demo test with Claude.ai and ChatGPT recorded. - [What 1.10.1 adds: playbooks, refusals and an origin check](https://agentabilitieswp.com/blog/whats-new-in-1-10-1-playbooks-and-clearer-refusals/): the Playbooks tab and its five starters, the two playbook tools, refusals that say why (paused connection, blocked address, rate limit, unknown ACF field), the origin check on the MCP endpoint, and the catalog moving from 182 to 184. - [Give an AI agent step-by-step instructions for WordPress](https://agentabilitieswp.com/blog/give-an-ai-agent-step-by-step-instructions-for-your-wordpress-site/): how to write a routine once as a playbook, what the agent receives when it loads one, and what a playbook cannot do. - [What 1.10.0 adds: control for each connection and three block tools](https://agentabilitieswp.com/blog/whats-new-in-1-10-0-control-each-connection/): pause, end-date or rate-limit each OAuth connection, access presets, Pause all access, and three block tools for agents. - [Pause, limit or expire an AI agent connection on WordPress](https://agentabilitieswp.com/blog/pause-limit-or-expire-an-ai-agent-connection-on-wordpress/): each per-connection control from 1.10.0, and what the agent sees while its connection is paused. - [Let an AI agent learn your site's blocks before it edits a page](https://agentabilitieswp.com/blog/let-an-ai-agent-learn-your-sites-blocks-before-it-edits-a-page/): the 1.10.0 tools that list block types, read a block's attributes and ACF fields, and replace one block. - [What 1.9.0 adds: a searchable Activity Log and alerts](https://agentabilitieswp.com/blog/whats-new-in-1-9-0-a-searchable-activity-log-and-alerts/): search and filters by user, connection, tool, plugin, date and text, the connection, IP and duration columns, refusal reasons that stay in the admin log, a 7-day summary and Live view, a CSV export that follows the filters, and alerts that stay off until you turn them on. - [Email alerts for risky AI agent actions on WordPress](https://agentabilitieswp.com/blog/get-an-email-when-an-ai-agent-does-something-risky-on-wordpress/): setting up the 1.9.0 alerts (admin notice, email right away or a daily email) for new apps, high-risk runs, denied-call bursts and switch changes, with the per-event email limit and the WP-Cron caveat. - [Claude: structured content does not match the tool's output schema](https://agentabilitieswp.com/blog/claude-says-content-does-not-match-the-tools-output-schema/): why Claude rejected bridged WooCommerce orders and products before 1.8.1, and the fix: update, then use Refresh tools list in Claude. - [How to find what an AI agent did on your WordPress site](https://agentabilitieswp.com/blog/how-to-find-what-an-ai-agent-did-on-your-wordpress-site/): use the Activity Log to filter by connection, tool and date, read denied rows and their reasons, and export a CSV. - [Review what your AI agent wrote before it goes live](https://agentabilitieswp.com/blog/review-what-your-ai-agent-wrote-before-it-goes-live/): what Force draft on create holds (new posts and publish requests) and what it does not (edits to posts already live), the Needs review card, and the Contributor setup that covers live edits. - [Why ChatGPT or Claude can't connect to your WordPress site](https://agentabilitieswp.com/blog/why-chatgpt-or-claude-cant-connect-to-your-wordpress-site/): a symptom table and three steps, starting with the Connection Doctor, then a CDN or firewall blocking the /mcp path, then a self-signed certificate. - [What we fixed: a switch that would not turn off](https://agentabilitieswp.com/blog/what-we-fixed-a-switch-that-would-not-turn-off/): on sites with a persistent object cache, read-only mode and the high-risk WooCommerce lock could refuse to turn off because the cache kept serving the deleted option value, and the activity log recorded the change as done anyway. 1.7.3 makes both switches read the stored value back and report an error naming the cache when a change does not take. - [What 1.7.4 adds: the catalog moves from 153 to 179](https://agentabilitieswp.com/blog/what-1-7-4-adds-the-catalog-moves-from-153-to-179/): the first release since launch that changed the ability catalog, from 153 to 179 (85 core, 94 integration). Five new integrations (The Events Calendar 13, Event Tickets 3 and read-only, Slim SEO 2, Avada 2, GeoDirectory 4 and off even when its host plugin is active), two new core abilities (a dry-run-by-default sitewide replace capped at 50 posts, and an SSRF-hardened media upload from a URL), a per-role and per-connection allowlist that can only narrow, and the OAuth paths that treated a failed database read as permission granted. - [A write that changed nothing now refuses by name](https://agentabilitieswp.com/blog/the-page-builder-guard-a-write-that-changed-nothing-now-refuses/): a write aimed at an Elementor, Divi, Beaver Builder, or Avada page used to report success and leave the visible page unchanged, because those builders keep their layout outside post_content. Since 1.7.4 the call is refused with a 409 naming the owning builder. It is a guard, not page builder support: the plugin gained no ability to read or edit those layouts, and OptimizePress is not detected yet. - [Absent and unknown are not the same answer](https://agentabilitieswp.com/blog/absent-and-unknown-are-not-the-same-answer/): WordPress returns the same value for "there is nothing here" and "the query failed", and 1.7.5 carries nine review rounds against that one shape. A failed consent lookup could read as consent granted and skip the consent screen; a failed read during an OAuth migration could switch OAuth back on for a site that had turned it off, live in 1.7.2 through 1.7.4; and refresh-token reuse detection reported a revoked chain while a concurrently minted successor survived, fixed by deactivating each node before reading its children so the InnoDB row lock serializes it. - [What 1.7.5 adds: picking abilities instead of typing them](https://agentabilitieswp.com/blog/what-1-7-5-adds-picking-abilities-instead-of-typing-them/): the per-role and per-connection allowlist added in 1.7.4 replaces its free-text slug box with a searchable checkbox picker, plus real role and client selects, a Client ID copy control, paginated Connection tables, and a cross-sub-tab abilities search. The stored data and the narrowing rules are unchanged; only the input control changed. The catalog stays at 179. - [What 1.7.6 changes: an OAuth token now works on the MCP route only](https://agentabilitieswp.com/blog/whats-new-in-1-7-6-an-oauth-token-now-works-on-the-mcp-route-only/): before 1.7.6 an OAuth token issued for the MCP endpoint also signed its holder in on other REST routes, admin-ajax.php, admin-post.php, the comment form, the authorize screen and the discovery URLs, because the plugin decided from the request URL before WordPress had routed it. From 1.7.6 it works on the MCP route only. The same release verifies meta, SEO, media, WooCommerce and Events Calendar writes against the database and reports written, unchanged, refused or a failed read, and reads read-only mode, the high-risk lock, ability switches and the allowlist from the database on every MCP request, keeping the stricter setting when a cached copy disagrees. - [What 1.7.7 changes: one Save changes button on the Abilities tab](https://agentabilitieswp.com/blog/whats-new-in-1-7-7-one-save-button-on-the-abilities-tab/): the Abilities tab had a main Save and four more buttons on its cards, and pressing only the main one lost edits to exposed meta keys and content types. 1.7.7 has one Save changes button, names a section that fails to save and keeps your edits, holds back sections that would widen access once another fails, and adds an Activity Log row when an exposed list changes, giving the list name and counts only. - [What 1.7.8 changes: bridged tools now show up in the tool list](https://agentabilitieswp.com/blog/whats-new-in-1-7-8-bridged-tools-now-show-up-in-the-tool-list/): 37 bridged tools from eight plugins worked when called but never appeared in the tool list. They now list for users who can edit posts, the source plugin still decides each call, and the aafm_bridge_discovery_capability filter raises the floor. Bridged errors name the source namespace, the Activity Log keeps a bridged plugin's short error code behind a foreign: prefix and never its message, and three bridged result bugs (ACF delete, plain-object results, exception text on WordPress 7.0 and later) are fixed. ## WordPress core coverage 90 abilities across: Posts & Pages, Terms & Taxonomies, Comments, Media, Post Meta (allowlisted keys only), Users (capability-limited), Site structure & menus, Revision history, Blocks & Templates, Limited settings & site health, Site-wide search across every post type. Reads plus guarded writes. Trash post and Trash page move content to the Trash, where it can be restored, but most Delete tools, such as Delete post, Delete page and Delete media, remove it permanently. The last administrator can never be removed. ## Clients that connect today ChatGPT, Claude.ai, Claude Desktop, Manus, Claude Code, Cursor, VS Code, Windsurf, Gemini CLI. Most connect directly or through the open-source mcp-remote bridge that runs on your own machine. ChatGPT, Claude.ai, Claude Desktop, and Manus connect by adding your endpoint as a custom connector over OAuth (ChatGPT adds it from Plugins as a custom MCP server; per OpenAI, full MCP with write tools is on Business and Enterprise/Edu plans, and Pro connects with read and fetch tools only). Claude Desktop shares the Claude.ai connector rather than running a local bridge, so a connector is dialled from the AI vendor's side and your site must answer on the public internet for those four. The Gemini app is NOT supported yet: it wants a streamable HTTP or SSE connector the underlying adapter does not serve natively yet. Each working client has a short setup guide: - [Connect ChatGPT to WordPress over MCP](https://agentabilitieswp.com/clients/chatgpt/) - [Connect Claude.ai to WordPress over MCP](https://agentabilitieswp.com/clients/claude-ai/) - [Connect Claude Desktop to WordPress over MCP](https://agentabilitieswp.com/clients/claude-desktop/) - [Connect Manus to WordPress over MCP](https://agentabilitieswp.com/clients/manus/) - [Connect Claude Code to WordPress over MCP](https://agentabilitieswp.com/clients/claude-code/) - [Connect Cursor to WordPress over MCP](https://agentabilitieswp.com/clients/cursor/) - [Connect VS Code to WordPress over MCP](https://agentabilitieswp.com/clients/vs-code/) - [Connect Windsurf to WordPress over MCP](https://agentabilitieswp.com/clients/windsurf/) - [Connect Gemini CLI to WordPress over MCP](https://agentabilitieswp.com/clients/gemini-cli/) ## Links - [Download on WordPress.org](https://wordpress.org/plugins/agent-abilities-for-mcp/) - [Source on GitHub](https://github.com/unaibamir/agent-abilities-for-mcp) - [Listed in the official MCP registry](https://registry.modelcontextprotocol.io/?q=io.github.unaibamir/agent-abilities-for-mcp): published as io.github.unaibamir/agent-abilities-for-mcp, where the current version is the entry marked isLatest. Being listed there is not a review or a rating.