You ask Claude to draft three blog posts and tidy up last month’s categories. To do that, it needs to get into your WordPress site. The quick way is to hand over an administrator login or an application password, and that gives the AI the same power you have: it can touch plugins, users and settings, publish anything, and delete anything.
A safer way is to give the AI its own identity on your site, with a short list of things it is allowed to do and everything else refused. It works through that agent, never as you. If something goes wrong, you switch the agent off without changing your own password.
That is how Zorem AI Agent Connector works, and the free version is now on WordPress.org. This guide shows how to connect Claude, ChatGPT or Codex to your site with it, how to check the connection, and what the free plugin does not do.
Why an admin login is the wrong key to give an AI
AI apps reach external systems through MCP (Model Context Protocol), an open standard that lets an app call tools on another system. A WordPress MCP server lists the tools your site offers, and the AI picks from that list.
What matters is whose permissions sit behind those tools. If the connection runs on your admin account, the tool list is limited only by what the connector exposes, and the AI acts with your full authority. A misunderstood instruction such as “clean up old drafts” then has nothing in its way.
An agent model puts a layer in between. The AI connects to the agent’s own MCP address, sees only the tools that agent has been granted, and has every other request refused. Three defaults make the difference in practice:
- New posts and pages are saved as drafts unless the agent specifically asks to publish.
- Deletion only moves content to the trash, and even that is off until you switch destructive tools on.
- The agent can never change its own permissions, manage users, install plugins or edit site code. The user list is read-only.
What you need before you start
- WordPress 6.9 or higher and PHP 8.0 or higher.
- HTTPS on your site if you want to connect by sign-in. AI apps only sign in to HTTPS servers. Without HTTPS, use the token method instead.
- An administrator account, to approve the connection.
WooCommerce is not required. The plugin is built on the WordPress Abilities API and includes the official MCP Adapter, so there is nothing else to install.
How to connect an AI agent to WordPress
1. Install the plugin
Go to Plugins > Add New, search for the free WordPress MCP plugin Zorem AI Agent Connector, click Install Now, then Activate. The plugin then sets up an agent called Admin Assistant the next time you open an admin page. There is no setup wizard to work through.
2. Copy the agent’s address
Open AI Agent Connector > Agent > Connection. Copy the agent’s MCP address. It looks like https://your-site.com/wp-json/ai-agent-os/v1/mcp/agent/…
3. Add it to Claude
In claude.ai or the Claude desktop app, open Settings > Connectors > Add custom connector. Give it a name, such as “Admin Assistant”, and paste the address into the remote MCP server URL field. Leave the OAuth client ID and secret empty. Click Add, then Connect.
Or add it to ChatGPT
Open Settings > Plugins, scroll to the bottom and turn on Developer mode. You have to do this first. Then go back, click Browse plugins, and click the + button beside the search box. Enter a name, keep Connection set to Server URL, paste the address, keep Authentication on OAuth, tick the acknowledgement and click Create. In a chat, open the + menu and select the plugin to use it.
Or connect Codex with a token
Codex can’t sign in, so it needs a token. On the agent’s page, issue a token from the Token card. It is shown once and stored only as a hash, so copy it straight away. Set it as an environment variable named ZOREM_ADMIN_ASSISTANT_TOKEN, then run the command shown in the Connection section:
codex mcp add admin-assistant --url <agent-address> --bearer-token-env-var ZOREM_ADMIN_ASSISTANT_TOKEN
Cursor, VS Code, Claude Code and other MCP apps connect the same way: add a remote MCP server, paste the address, and sign in if the app supports it. If it doesn’t, send the token as an Authorization: Bearer header.
4. Approve the connection
Your site shows a connection request screen. Before you click Approve and connect, check three things: the app’s name, the agent it will act as, and the list of tools. Destructive tools are marked. If you didn’t start the connection yourself, click Deny. Approving doesn’t change any permissions; it only lets this app use the agent as it is currently set up.
5. Narrow the permissions
The free plugin includes 43 WordPress tools. They cover posts, pages, categories and tags, comments, the media library, a read-only user list, navigation menus, custom post types, and site tools such as statistics, post meta, search and replace, and cache clearing.
On the Permissions tab, set each tool to None, Read or Write. Read-only tools offer only None and Read. Each section heading has Grant all, All read and Revoke buttons for bulk changes. Nothing is saved until you click Save permissions.
Check this tab before you give the AI any real work. Out of the box, the Admin Assistant has Read on every read-only tool and Write on everything else. That’s convenient for a first test, but it’s broader than most sites need. Revoke whole sections you won’t use, such as menus or custom post types. Pay particular attention to search and replace: it is rated Write, not Destructive, so the destructive switch doesn’t block it, and it can rewrite text across your whole database. Set it to None unless you have a specific job for it.
6. Decide on deletion and personal data
Under Access on the Connection tab are two switches only an administrator can change:
- Allow destructive abilities: off by default. Until it’s on, tools that move content to the trash are refused. Even when it’s on, nothing is permanently deleted.
- Mask customer personal data: off by default for the Admin Assistant. Turn it on and names, emails, phone numbers and addresses reach the AI app as masked placeholders.
Click Save changes. New settings apply from the agent’s next request. The access and safety documentation covers each switch in more detail.
How to confirm the connection works
- Check the tools appear. In Claude, the agent’s tools appear in your chats once the connector is connected. In ChatGPT, they’re available after you select the plugin from the + menu.
- Ask a read-only question first, such as “Which agent am I connected as, and what can you do on this site?” The core “Who am I” tool answers this without changing anything.
- Try one small write, such as “Draft a post titled Test”. Then check Posts: it should be there as a draft, not published.
- Open the Status screen if anything fails. It shows versions, database tables, MCP Adapter status and any issues, with diagnostics you can copy into a support ticket.
The most common failure is easy to miss. If sign-in appears to succeed but every call is refused (Claude reports that the connector “returned an error when connecting”), your web server is probably removing the Authorization header. The Connection section warns you about this and shows the line to add to your server configuration.
What the free version does not do
- There is one agent. Every app you connect signs in as the Admin Assistant with the same permissions. Separate agents for separate apps or tasks are part of the Pro version, AI Agent OS.
- There is no activity log. The plugin doesn’t record the agent’s calls. The Status screen keeps only the most recent refused connection. Changes show up where WordPress normally keeps them, such as post revisions, comments and menus. A log of every call and refusal is in AI Agent OS.
- The 43 tools are WordPress tools, not WooCommerce tools. If another active plugin registers abilities through the WordPress Abilities API, its tools appear on the Permissions tab with access set to None until you grant them.
- It won’t stop a publish you asked for. Drafts are the default, but if you tell the agent to publish and it has Write access, the post goes live.
- Personal data isn’t masked by default. The Admin Assistant acts for the administrator and sees real data until you turn masking on.
On privacy, the plugin itself sends nothing to Zorem: no tracking, no telemetry, no calls to Zorem servers. The only data that leaves your site is what the connected AI app receives through the agent. That data is then handled under the AI provider’s own terms, so masking is worth turning on if your posts or comments contain customer details.
Frequently asked questions
Is Zorem AI Agent Connector really free?
Yes. It’s free and open source on WordPress.org, with all 43 WordPress tools, per-tool permissions, sign-in and token connections, data masking and the destructive switch. AI Agent OS is the paid version. It adds separate agents and an activity log for teams that need to see every call an agent made.
Can I connect Claude and ChatGPT at the same time?
Yes. You can connect several apps at once, and each one signs in as the Admin Assistant. They all share the same permissions, which you manage from one place. To give different apps different permissions, you need separate agents, which is a Pro feature.
Can the AI delete my content?
Only if you turn on “Allow destructive abilities”, and even then it can only move content to the trash. Permanent deletion isn’t available to the agent. The switch is off by default, and only a site administrator can change it.
Why does sign-in say my site isn’t HTTPS?
AI apps only complete sign-in with servers that use HTTPS. If your site doesn’t have HTTPS yet, issue a token on the agent’s Connection tab and connect with that. For a live site, enabling HTTPS is the better long-term fix.
How do I disconnect an app?
Remove the connector inside the app, or click Revoke this agent on the Connection tab. Revoking stops the token and ends every sign-in straight away. Permissions are kept, so if you reactivate the agent it comes back as it was, although apps that signed in will need to be approved again.
What happens if I click New address?
The old address stops working, and every app connected to it has to be added again. Use it if you think the address has been shared somewhere it shouldn’t be. Issuing a new token is different: it only affects apps that use the token, not apps that signed in.
Start with read access and widen it later
Connecting an AI app to WordPress takes a few minutes. Deciding how much it’s allowed to do is the part worth taking slowly. Install the plugin, trim the default permissions, connect one app, leave the destructive switch off, and run a week of real tasks before you grant more. Everything the agent changes shows up in your normal post revisions, so you can review its work the same way you’d review a new writer’s.
The connection guide in the docs covers each app in more detail, including troubleshooting..

