Connecting a client
Your AI client connects in to your site over MCP. The plugin never calls out to any AI provider, so there is no model API key to add and nothing extra to pay for. There are two ways to connect.
The two ways to connect
Section titled “The two ways to connect”Approve the agent once in the browser. Paste your site’s MCP endpoint into your client, approve the connection, and the agent acts as the account that approved it. There is no secret to store. Approve while logged in as the account you want the agent to be, because the connection takes that account’s capabilities. When you approve, you can also pick an access preset (Read only, Content editor, Shop manager or Custom) that narrows what this connection reaches.
Point a dedicated low-privilege user at the endpoint. Create the agent user the Connection tab offers, generate an Application Password for it, and connect with that. A guided screen builds the client config and checks the endpoint for you. Read the caution below before you generate one: an Application Password is a whole-site credential and this plugin cannot narrow it.
Clients that connect today
Section titled “Clients that connect today”ChatGPT, Claude.ai, Claude Desktop, Manus, Claude Code, Cursor, VS Code,
Windsurf, and Gemini CLI all work today, some directly and some through the
open-source mcp-remote bridge.
- Hosted apps: ChatGPT, Claude.ai, Claude Desktop, and Manus. Each connects by adding your site’s endpoint and approving it once over OAuth. In ChatGPT on the web, open Plugins, select the plus button, then Add custom MCP server, and enter your endpoint with OAuth as the authentication. You do not need to turn on Developer Mode first. OpenAI’s help center lists full MCP, write tools included, for Business and Enterprise/Edu plans, while Pro can connect with read and fetch tools only. On Claude.ai you add a custom connector under Customize, then Connectors, and a Free plan can add one. Claude Desktop shares that same connector, so one you add in either place shows up in both. On Manus you add your site as a custom MCP connector using your endpoint URL. ChatGPT and Claude register themselves with your site during that first approval, which needs dynamic client registration switched on. It is on by default from 1.7.2, including on sites that never touched the setting. Since 1.10.0 its switch sits on the Connection tab, under OAuth, in the Token lifetimes and client registration card. Connectors are dialled from the AI vendor’s side rather than from your own machine, so your site has to answer on the public internet. A site on localhost or behind a VPN will not connect this way, and Claude Desktop is no exception despite running on your own computer.
- After you change the enabled abilities: ChatGPT keeps the tool list it saw when it connected. Open Settings, then Integrations, then Plugins, pick the app and choose Refresh, or it will not see the tools you just switched on or off. In Claude, start a new chat or reconnect the connector.
- AI code editors and IDEs: Claude Code, Cursor, VS Code, and Windsurf. Since 1.12.0 the Connection tab’s Cursor and VS Code quickstarts have Add to Cursor and Add to VS Code buttons. They pass the editor your endpoint and a server name, never a credential, and the editor then runs the OAuth sign-in itself.
- Command line: Gemini CLI.
- AI agent frameworks: any MCP-compatible framework can call your enabled abilities as tools.
The bridge, for clients that cannot connect directly
Section titled “The bridge, for clients that cannot connect directly”Clients that cannot open a remote MCP connection on their own use the
open-source mcp-remote or @automattic/mcp-wordpress-remote bridge. It runs on
your own machine and talks only to your site and your local AI client. Neither
bridge is bundled with the plugin or run by it: you install and run it yourself.
The Gemini app no custom servers
Section titled “The Gemini app ”We have not found a way to add a custom MCP server in the Gemini app, so there are no steps for it here. Gemini CLI does connect: it runs on your own machine, like Claude Code, and the Connection tab has a ready-made quickstart for it. The clients listed above all work today, including ChatGPT as a custom MCP server.
Revoking an agent’s access
Section titled “Revoking an agent’s access”Connecting an agent is documented everywhere. Cutting one off usually is not. Uninstalling the plugin does not revoke an agent on its own. By default it keeps the plugin’s settings, activity log and OAuth data, and removes them only if you turned on “Delete data on uninstall” in Settings first. What it never removes is the dedicated agent user or any Application Password issued to it, because those are ordinary WordPress account credentials that live outside the plugin’s own data.
To actually cut an agent off, do one of these:
- Revoke its OAuth grant on the plugin’s Connection tab, under Active grants.
- Revoke its Application Password on the Connection tab’s Application Passwords tab, which lists the passwords on the agent users this plugin created, or delete it on the Users screen, under that user’s profile.
- Delete the agent user account outright, if it exists only to serve that agent.
Any of the three works before or after you remove the plugin. Do it first if the credential is the thing you want gone, since removing the plugin closes the door this plugin owns and leaves the WordPress credential sitting there.
If you only want a connection to stop for a while, pause it or give it an end date instead of revoking it, and use Pause all access to stop every MCP request at once. Controlling each connection covers both.
If your site moves to a new address, OAuth tokens issued for the old one can no longer connect. Since 1.12.0 the Connection tab says how many there are and offers an End old tokens button. Ending them removes no one’s approval, but each person who approved one of those apps has to remove the connector and add it again with the new address.
Next steps
Section titled “Next steps”- Controlling each connection: pause, end date, rate limit and access presets.
- Troubleshooting: OAuth discovery, Windows clients, local sites.
- Governance concepts: what the connected agent can and cannot do.