| 开发者 | unaibamir |
|---|---|
| 更新时间 | 2026年11月16日 06:50 |
| PHP版本: | 7.4 及以上 |
| WordPress版本: | 7.1 |
| 版权: | GPL-2.0-or-later |
| 版权网址: | 版权信息 |
mcp-remote, which runs on your own machine.
Disclaimer
The Model Context Protocol is an open specification that Anthropic originally developed. Claude, ChatGPT, Cursor, VS Code, Gemini and other product names belong to their owners, and Agent Abilities for MCP is a third-party plugin that isn't affiliated with, endorsed by or sponsored by any of them.
/wp-content/plugins/agent-abilities-for-mcp, or install it from Plugins, Add New in your admin, then activate it.More help: Documentation | Troubleshooting | Security | Support forum A WordPress site that AI apps can call to do things like list posts or update a product, with this plugin supplying the endpoint and the rules about who can use it and for what.
Follow the steps under Connect Claude to WordPress above, which add your endpoint as a custom connector under Customize, then Connectors. The one thing to add here is that Claude connects from Anthropic's cloud, so a site on localhost or behind a VPN won't work. The Claude guide has the full setup.
You don't need to turn on Developer Mode first, and the steps are under Connect ChatGPT to WordPress above (the ChatGPT guide walks through them too). ChatGPT reaches your endpoint over the public internet, so a site on localhost or behind a VPN won't connect, and because custom servers are a web feature you can't add one from the mobile app.
It can be, as long as you keep the access small: the AI works as one WordPress user you choose, this plugin's tools start off, and each call to them is checked against that user's capabilities and logged. I'd start with read-only mode or Force draft on create and widen from there once you've seen what it does. What ChatGPT or Claude does with the content it reads is up to that vendor, and text on your own site (a comment, say) can try to steer an AI, a risk called prompt injection that the post on indirect prompt injection covers.
The code is GPL-licensed and public on GitHub, the plugin makes no calls to AI providers and sends no telemetry, and it stores OAuth credentials as SHA-256 hashes in its own tables. The security page lists the controls and says how to report a problem.
Not unless you connect it as an administrator, which I wouldn't. It gets the role of the user it connects as: for OAuth that's whoever approves the connection, and Roles that can connect on the Settings tab decides who may approve, with a new install allowing Administrators only. The Application Password route uses a Subscriber account the Connection tab creates, which you can raise to Editor or Author. An Application Password works across the whole REST API and this plugin's limits apply only to calls through its MCP endpoint, whereas a token issued over OAuth works on that endpoint alone.
Yes, through the user's role. An Editor can write and publish any post but can't touch site settings or install plugins, an Author is confined to its own posts, and a Contributor writes without publishing. The Content editor preset narrows it further to posts, pages, media, categories and tags, and because a new install lets only Administrators approve a connection, add the role under Roles that can connect first or approve as an administrator and pick the preset.
Flip Pause all access on the Connection tab, the emergency stop for every connected app and Application Password. Nothing is deleted and Resume all access turns everything back on, but since Application Passwords keep working on the rest of the REST API you should revoke one if it has to go.
Yes. Each approved OAuth connection on the Connection tab has Pause next to Revoke, and a paused connection keeps its sign-in and can still list tools while every tool call comes back with a message that the connection is paused. Resume it and nobody has to approve it again. The same person's Application Password is unaffected, since it has no pause of its own.
For an OAuth connection, yes. Set an end date on its row on the Connection tab and the connection stops at the end of that day and stays refused until a site administrator renews it, and its old tokens never work again. An Application Password has no end date, so revoke it when the job is done.
Turn on read-only mode in Settings and this plugin loads no write tool, whatever else is ticked, including write abilities you bridged from other plugins. Search and read keep working, and your selections come back when you switch it off. To limit a single connection instead, pick the Read only preset when you approve it.
With Force draft on create switched on in Settings, anything the AI creates and any request to publish is saved as a draft, which covers posts, pages, custom content items, reusable blocks and WooCommerce products. Product variations save as disabled instead, because a draft variation disappears from WooCommerce's editor. Edits to a post that's already live still go live, so if every change should be reviewed, connect the AI as a Contributor, who can't edit published posts. The drafts wait on the Needs review card on the Dashboard.
I'd leave the delete tools off, which is how they ship. Trash post and Trash page move content to the Trash where you can restore it, but most of the tools named Delete, such as Delete post, Delete page and Delete media, remove it permanently. Read-only mode takes away every write and delete tool of this plugin's at once and the Content editor preset has no deletes in it, and the Activity Log shows which tool ran on which item though it keeps no copy of deleted content.
Open the Activity Log tab (the guide to finding what an agent did walks through it). Each row gives the time, user, connection, source IP, tool, run time and a short note on what the call touched, such as a post ID or a meta key name, and a 7-day summary sits at the top. Export CSV downloads whatever your filters show. The log records which tool ran on which item and never stores the text the AI wrote.
Yes, once you set it. Enter a number under Rate limit (per minute) on the Settings tab and each agent user can make that many calls a minute, shared across all its connections, with 0 turning the limit off. One OAuth connection can have a lower limit of its own on the Connection tab, in which case its calls count against both, and calls over a cap are denied and logged.
Yes. Enter addresses or CIDR ranges, one per line, under IP allowlist on the Settings tab and requests from anywhere else are refused before they sign in. Behind Cloudflare or another proxy, switch on Real visitor address behind a proxy or CDN first, because otherwise the plugin sees the proxy's address and the allowlist can lock your agents out.
A preset caps what one OAuth connection can reach, and you pick it when you approve an app or in Quick Connect and can change it later on the connection's row.
Step-by-step instructions you write on the Playbooks tab, for example how to check a draft before it goes out. A site can keep 50 of them, with steps up to 8,000 characters, and you can add your own or start from one of the five starters. Once List playbooks and Get playbook are switched on, the AI sees your Active playbooks at the start of a session and loads one when a task fits, and the list refreshes when the client reconnects.
Yes. Put the plugin on every site and add each endpoint as a separate connector, and name the connectors after the sites so you can tell them apart in the chat.
With ACF active, the ACF tools list field groups and read and write field values on posts, terms and users, sanitizing each value for its field type, and a write that names a field ACF doesn't know is refused with an error that names it. For layout, the AI can list block types and read one block's attributes, plus field names and an example for an ACF block. Replace block then swaps one block in a post, found by its position, and leaves the others alone, although WordPress may still sanitize the whole post on save for users who can't post unfiltered HTML. Questionable block markup saves with a warning unless you turn on Strict block validation in Settings.
A page builder owns its page's layout, so an edit to the post text could report success and change nothing. Before any write, update-post, update-page, replace-in-post and the other general content tools check whether a page builder owns the post, and if one does they refuse with an error that names the builder. Elementor, Divi, Beaver Builder, Avada, Visual Composer, Oxygen, Breakdance, Brizy, Bricks and WPBakery are covered. The one write they let through is an edit that only sets the excerpt, which leaves the stored content as it was, and even that is refused when the plugin can't tell which builder owns the post. Pages built with the other eight builders are edited in the builder itself. Avada and WPBakery Page Builder pages have editing tools of their own: they read a page as an outline of elements, change an element's settings or text, add, move and remove elements, and place a saved template or Library item. Each write is refused, with nothing saved, if saving the page as your user would change anything else on it. Restore revision works on these pages too, and when WordPress keeps a revision the editing tool returns it so the page can be brought back. Remove element is refused when no revision could bring the element back, which means revisions are off, the site keeps only one revision, or the target is an Avada form or off canvas panel. All of these tools start off, and the writes are in no preset except Custom.
Text, links, images, colours and spacing work for anyone who can edit the page. Code blocks, raw HTML and JS, Design Options, page CSS and JS, and settings that can carry script need a user who may post unfiltered HTML, and so do square brackets, percent-encoded brackets and data-* attributes in body text. The site-wide tools (WPBakery settings, Avada Global Options, Avada layouts and their sections, Avada Slider settings and both Avada Studio tools) need a user who can manage options, and so does editing the elements of an Avada layout section. Neither builder's Role Manager is applied, so the agent gets what its WordPress user can do and no more. With Templatera active, its templates are listed in place of WPBakery's saved templates and can be copied into a page or placed by reference, and a Templatera template can be edited like a WPBakery page. With Avada Core active, an agent can change an Avada Slider's settings, and Avada page options also work on Avada Slider and Elastic Slider slides. The settings tools return the previous values, which is the way back. The Avada Studio tools start switched off. The import has its own switch, which no Enable all button ticks, and the list tool is refused until the import is on. When an agent uses them, Avada downloads the catalogue and items from avada.studio under your site's Avada registration; this plugin sends nothing in that exchange and only checks with Avada that Studio is available. Avada's terms and privacy policy apply. Refused on both builders: an element type the builder's own element map doesn't list, and a stale element index or page fingerprint. A global Avada Library element is edited in the Library, not on the page that shows it. Adding an element to a WPBakery flexbox or grid container item, or moving one into it, isn't supported yet, though elements already inside one can be reordered, updated and removed.
With WooCommerce active, 52 abilities cover products and variations, orders with notes and refunds, customers, coupons, shipping, tax, payment gateways and reports, and because they reach customer data an admin notice flags them. The tools that move money or grant authority stay locked until you unlock them in Settings, and the Shop manager preset leaves them out.
No, the official MCP Adapter is bundled, and if the standalone Adapter plugin is active too, this plugin uses whichever copy loads first. It builds on the WordPress 6.9 Abilities API. The Adapter also has a built-in default server that other plugins can publish tools to, and those tools sit outside this plugin's controls: its read-only mode, allowlists, IP allowlist, rate limit and audit log don't apply to them, though Pause all access does. Under Settings, Allow the MCP Adapter's default server starts on Automatic, which runs it only when another active plugin publishes tools there. Choose Always off to keep it off, or Always on if you ran the standalone Adapter only to reach WordPress core's public abilities, since those don't count as a publisher. This plugin's own tools are never on that server.
That notice comes from a copy of the MCP Adapter that runs its own loader and finds the adapter already loaded by a different copy. This plugin bundles adapter 0.7.1 and loads it without that loader, so its own copy never prints the notice. When the standalone MCP Adapter plugin is active too, this plugin asks it to step aside quietly, unless your site already sets the WP_MCP_AUTOLOAD constant. If the notice still shows, another plugin ships its own copy of the adapter. The Connection Doctor report on the Connection tab shows which copy is running. If it says bundled, this plugin's copy is in use and the notice comes from the other plugin's copy, so update or remove the plugin that bundles its own MCP Adapter.
Because this plugin's tools are all off until you switch them on, it's a poor fit if you want everything available from day one. In exchange the AI acts as a user you choose, each OAuth connection can be paused or given an end date, and the plugin builds on the Abilities API and the MCP Adapter instead of running a server of its own.
The plugin never contacts an AI provider and has no telemetry, but your AI client does receive whatever a tool returns, because that is how it reads your site, and what it does with that is covered by the vendor's terms and your plan. The only request the plugin makes to an outside address is the upload-media-from-url tool, which stays off until you enable it, and External Services above has the details. If you switch on the Avada Studio tools, Avada itself fetches the Studio catalogue and items from Avada's servers under your Avada registration, and this plugin adds nothing to those requests.
It's free on WordPress.org. There's no paid tier, the plugin adds no usage limits and it needs no API key. You need WordPress 6.9 or newer, where the Abilities API arrived, and PHP 7.4 or newer. You can get connected without writing code, since the Connection tab gives you the endpoint to paste into your AI client, and developers get filters and a WP-CLI command that are documented on GitHub.
Click Check my connection on the Connection tab and the Connection Doctor runs its checks from your own server, then names the fix, whether that's a host stripping the Authorization header, a firewall emptying the JSON body, a redirect or a blocked discovery document. The nginx rule and the Cloudflare and WP Engine steps are in the troubleshooting guide, and server rules that block /.well-known/ or CDN rules such as Cloudflare's AI bot blocking are common causes.
That error means ChatGPT never got the plugin's discovery document, so it couldn't see the PKCE support listed there. The usual cause is a firewall or CDN rule blocking ChatGPT, or a host serving .well-known from disk before WordPress sees the request. Test from outside your network with curl -sS -A "ChatGPT-User/1.0" https://your-site/.well-known/oauth-authorization-server (on a subfolder install, add the folder: https://your-site/blog/.well-known/oauth-authorization-server), and if the JSON doesn't come back, allow that path for ChatGPT.
The AI client may be showing an old tool list. To refresh it in ChatGPT, open Settings, then Integrations, then Plugins, pick the app and choose Refresh, and in Claude start a new chat or reconnect the connector. If that isn't it, the Why is my tool missing? card on the Dashboard lists each of this plugin's own tools that you switched on but agents can't see, with the reason, such as read-only mode, the high-risk lock or an inactive plugin.
The MCP endpoint refuses requests sent from a web page on another site, as the MCP specification requires, while desktop apps, command-line tools and the hosted ChatGPT and Claude apps are unaffected. Developers can allow a browser client they trust with the aafm_mcp_allowed_origins filter.
Under Settings, Notifications, pick Admin notice only, Email alerts as they happen or One email a day, then tick the events you care about, such as a new app being authorized or a burst of denied calls. Nothing is sent until you do. The safety controls guide lists every event and the sending limits, and the emails themselves link to the Activity Log and never include the content of a call, a user name or an IP address. Where WP-Cron never runs, an alert can arrive an hour or more late.
Yes, for edits made with 25 of this plugin's tools: posts, pages and custom post types, creates, trashes, block edits, terms, custom fields (post and term meta), featured images and SEO fields. The Changes tab keeps up to 30 days and reverts a change if the item hasn't been edited since, and it refuses a revert that would overwrite a newer edit. It lists field names, never contents, though the earlier values are stored in a table in your site's database, and when the history is full the oldest values go first. No other tool of this plugin's can be undone, including permanent deletes, orders, refunds, user accounts, file uploads, WooCommerce products, ACF field values, menus, comments and events, and each of them says so in its response. That includes the WPBakery and Avada tools, but a page edit keeps a WordPress revision when revisions are on, and the tool returns it for restore-revision; their settings tools return the previous values instead, and an imported Avada Studio item is removed by deleting it in Avada's Library. Tools from other plugins aren't recorded either. Notifications, webhooks and anything already read are not recalled. Some plugins, GeoDirectory among them, turn revisions off for the whole site; a WP_POST_REVISIONS line in wp-config.php turns them back on. Two optional tools, list-changes and undo-change, let an agent undo its own changes from the last 24 hours, and both start off. History records only while the plugin is active, can be switched off in Settings and is removed on uninstall. WordPress's export and erase personal data tools cover it.
Not by itself. Deleting the plugin leaves the dedicated agent user and its Application Passwords in place because they're ordinary WordPress accounts, and by default it keeps your settings, activity log and OAuth data too, so a reinstall picks up where you left off. Switch on Delete data on uninstall in Settings first if you want that data removed (change history is always removed, whatever that switch says), and to cut an agent off, revoke its OAuth connection on the Connection tab or delete its Application Password or user on the Users screen.
Use the Report a vulnerability option on the Security tab of the GitHub repository so the report stays private until there's a fix. The security policy lists the other contact route.