AI agents on a WordPress site: a team setup routine
A team that lets AI assistants near a WordPress site needs a routine more than it needs a long rulebook: one person who connects each app, a short list of what that app may do, and a named person who looks at what it did. BetterShield › Agents has the settings for all three, and this post puts them in the order a team can follow.
The rules behind each setting are in MCP security best practices for a WordPress site, the per-app tool choice is in WordPress MCP permissions: limit each AI app’s tools, and the change side is in WordPress MCP write access: approvals, limits and undo. This post adds the team layer: who does what, how often, and what happens when a person or an app leaves.
Quick summary
- Name three jobs before anyone connects an assistant: who connects it, who decides how much it may do, and who reviews what it did. One person can hold all three on a small team, but write the names down.
- Allow AI assistant connections opens the connection and turns on reading with it. Fixing is a second, separate switch, Let AI assistants fix things, and it starts off.
- Give each app its own connection and its own Tools selection, so one person’s assistant does not share a door with another’s.
- Look at Activity › Agent activity on a fixed day each week, starting with the refused calls. Rows are kept for 30 days in the free plugin, so a weekly look stays inside that window.
- When someone leaves, end their connection under Connected apps. A new WordPress password does not end a connection.
Why a team needs a routine
An assistant connects to the site as an account and can reach only as far as that account can. On a one-person site that is easy to keep in your head. On a team it is not: several people may connect several apps, each connection looks the same on the screen, and the question “who allowed that one?” comes up months later.
A routine answers it in advance. Each connection has an owner, a purpose and a review day, and the site keeps the record. We recommend writing the three names and the review day in whatever place your team already keeps its runbooks.
| Job | Who holds it | What they do |
|---|---|---|
| Connect | The person whose account the assistant will run as | Adds the app, tests it, limits its tools |
| Decide | Whoever owns the site | Chooses whether any app may make fixes |
| Review | A named person, with a back-up | Opens Agent activity each week and acts on what looks wrong |
Step 1: connect one app at a time
The person who connects an app should use an account that has only the access the work needs, because the assistant runs as that account. Then, on one site:
- Open BetterShield › Agents › Connect and turn on Allow AI assistant connections. It is off by default, and turning it on also turns on Let AI assistants see security info.
- Under Add BetterShield to your assistant, choose the app. For Claude or ChatGPT, copy the address, add it to the assistant and sign in when your browser asks. An editor or command-line tool uses Set Up an Editor and a connection key instead.
- If you create a key, leave Allow this key to make fixes unticked, so it can only read. The key is shown once, so treat the copied configuration as a secret and keep it out of screenshots and tickets.
- Press Test Connection under Connection diagnostics and read the actual result, including any failed check, not only the tick.
That gives the team a read-only assistant. We recommend living with that for a week before anyone asks for more: ask it which finding to fix first, and judge whether its answers are worth giving it more reach.
Step 2: limit what each app is offered
Under Connected apps, each app has a Tools list. Keep only the tools its job needs and select Save Tools. The app sees the saved selection the next time it reads its tool list, so ask it to refresh or reconnect before you test.
This is an extra boundary. It never grants anything that the site’s switches, the credential or the account forbid, and a limit on one app does not change another. That is why a separate connection for each app is the point of the routine: it is what lets you answer “what does this one need?” one app at a time.
Keep the tools your own workflow relies on. BetterShield 1.1.8 and later warn before a selection removes the reads that BetterShield Hub uses, so read that warning before you save.
Step 3: decide who may allow changes
Reading and changing are separate decisions, and on a team the second one belongs to the site’s owner. Let AI assistants fix things, on Agents › Permissions, is off by default. A connection made while it was off stays read-only, so reconnect the assistant if it should change the site.
Even then, the changes are sorted by size:
- Small fixes that can be put back apply directly, and each is logged.
- Heavier changes come as a plan, which says what would change and whether it can be put back. The assistant is told to show it to you and act only after you agree in the conversation, and the plan expires after 15 minutes.
- Some things never go through an assistant: making an account or credential, changing recovery options, weakening two-factor or alerting, lifting a lockout, removing log rows, or writing
wp-config.phpor.htaccess. Agents › Permissions lists them under What stays yours.
A plan from another caller, such as WP-CLI, waits under Agents › Requests, where Allow once allows that single change and Decline sets it aside. We recommend that the person who approves a plan is not only the person who asked for it, where your team is large enough to split the two. Review change requests from agents explains the screen.
Step 4: look after many sites from BetterShield Hub
A team that looks after several sites does not have to repeat step 1 on each one. BetterShield Hub is free and optional. Connect each site from its own BetterShield › Hub screen, then connect your assistant to Hub once from Agents › Connect in Hub. It can read the workspace sites that the connecting person can see.
Hub gives the person who owns the workspace the same kind of limits, per assistant, under Connected assistants:
| Limit | What it does | Default |
|---|---|---|
| Read only | Blocks any change through that connection | Off |
| Only these sites | Narrows it to the sites you choose, for example one client’s | Every site its person sees |
| Tool selection | Clears the tools unrelated to the assistant’s job | Every offered tool allowed |
Limits narrow the owner’s role and never grant access the person does not already have. We recommend starting an assistant read-only and giving it only the sites it needs. On each site, the Connections and safeguards card in Hub shows the connections the site reports and whether its read, change and activity-recording safeguards are available, and it tells you when its report is stale. An agent that cannot open a browser can use Hub’s connection instructions, which end with an approval you give in your own browser. BetterShield Hub assistants: connect, limit and review across sites covers that Hub layer in full.
Step 5: the weekly look
The review is the part teams skip, and it is the part that makes the rest worth doing. Record what agents do on this site is on by default, and it records calls to abilities from any plugin, refused calls included. It needs WordPress 7.1 or newer. Inputs are stored as a digest, never as their contents.
On the review day:
- Open BetterShield › Activity › Agent activity and set the outcome to Refused. A refused call is the clearest sign that an app tried something it was not allowed to do.
- Set Came through to MCP to see what assistants ran, and compare it with the people and apps you expect.
- Look for a call you cannot connect to a person. Use Connected apps and the account named on the row to find whose connection it came through.
- If your team reports to someone, run
wp bettershield agents --format=csvfor the matching rows, or add-for refused calls only. The screen itself has no export.- outcome=permission_denied
Filter and review agent activity lists every filter. A namespace in the list is a naming convention, not proof of who is behind an ability, so treat it as a clue and not as an answer.
Step 6: when a person or an app leaves
Connections belong to the people who made them, so the clean exit is part of the routine.
- Under Connected apps, choose Disconnect for the departing person’s app. It ends that app’s access.
- If they used a connection key, Replace Key makes a new one and the old one stops working at once. Disconnect All ends the key and every app that signed in through the browser, so use it when you cannot tell which is which.
- Changing the person’s WordPress password does not end a connection, so do both.
- Disconnecting stops future access. It does not undo anything the app already did, so check Agent activity for what it ran and reverse any change you do not want to keep.
In Hub, Disconnect on Connected assistants ends that connection the same way, and it cannot be restored, so connect the assistant again if you removed it by mistake.
What it can break, and how to undo
- A tool left out of an app’s list can stop a workflow that needs it. Open Tools for that app, tick the tool and select Save Tools.
- Turning off Allow AI assistant connections pauses every connection without revoking any of them. Turn it back on to resume.
- Turning off Let AI assistants fix things stops new changes at once, even for a connection that was allowed to make them. It does not undo a change already made, which you reverse separately.
- Disconnecting an app cannot be restored. Connect the app again if you did it by mistake.
Check it is working
Ask the assistant for a permitted read, such as the security score, then open Agent activity and confirm a completed row appears with the right door. If you can, also make one call you expect to be refused, and confirm that a refused row appears with a different outcome. Then press Test Connection again after any change to a key or a tool list.
Recommended settings
We recommend a read-only connection for every app to begin with, a Tools selection for each, fixes on only for a dedicated and reviewed connection, Record what agents do on this site left on, and a weekly look at Refused. For a team, write down who connects, who approves and who reviews. The Model Context Protocol is the open standard assistants use to call tools on a site like yours, and AI assistants and agents shows how the pieces fit together.
Frequently asked questions
Who on the team should connect an assistant?
The person whose account the assistant should run as. It can reach only as far as that account can, so connect it from an account that has the access the work needs and no more.
Can each person use their own assistant?
Yes. Give each app its own connection and its own Tools selection, so you can limit or end one without touching the others.
How often should we review agent activity?
We recommend once a week, on a fixed day. Rows are kept for 30 days in the free plugin, so a weekly look never lets anything age out unread.
Does a new password end an assistant’s access?
No. A new WordPress password does not end a connection. End it under Connected apps, or with Replace Key or Disconnect All for a connection key.
Do we need BetterShield Hub to do this?
No. Everything above works on one site without it. Hub is free and optional, and it helps when one assistant should read several sites with its own limits.
Conclusion
A team’s AI setup holds up when each connection has an owner, a short list of tools and a review day. Connect one app at a time, keep it read-only until there is a reason, look at the refused calls every week, and end connections when people move on.
BetterShield is free on WordPress.org. The docs for this setup are AI assistants: connect, approve and limit access and Agent activity.