Skip to main content
Every tool a remote AI app calls has a tier, and the Dashboard enforces it on every call. The connection’s approval policy (Standard, Strict or Hands-off) moves tools between tiers. The app can’t skip a tier: a call that arrives without the approval its tier needs doesn’t run.

The four tiers

A connection with View only access can call read tools only. Every other tier needs View and make changes.

Confirm: a preview in the chat

For a confirm-tier call, the Dashboard runs the tool’s preview where it has one, and otherwise lists what was asked. It returns CONFIRMATION_REQUIRED with the preview and a confirmation token. The app should show you the preview, wait for your reply, and call the tool again with the same arguments, user_confirmed: true and the token. The token is valid for 10 minutes and works once. Before it runs, the Dashboard checks again which sites the call affects. For tools that compare their preview, it also runs the preview again and compares the parts that describe the change. If either differs, it returns a new preview instead. The app sends that confirmation itself, so this tier proves the app passed the preview’s token back, not that a person read the preview. Only the Dashboard tier proves that.

Dashboard: your click on the approval page

For a Dashboard-tier call, the first call runs nothing. The Dashboard stores the request and gives the app an approval link. See The approval page.

Which tools are in which tier

These are the tiers under Standard. Tool names are the ones AI apps see. Tools that change Team Control roles and users are always Dashboard, except under Hands-off, where they are Confirm. Four core tools change tier with their arguments:
  • update_site_v1 is Dashboard when the call sets url, verify_certificate, ssl_version, http_user or http_pass; Confirm when it sets suspended; and Write otherwise.
  • add_client_v1 and update_client_v1 are Confirm when the call includes selected_sites, because that moves sites between clients, and Write otherwise.
  • update_backup_destination_v1 is Dashboard when the call sets minimum_remote_retention_days, because 0 removes the protection on older remote backups, and Confirm otherwise.

Backup calls that are refused

These calls are refused under every policy, and the message says what to do instead:
  • create_backup_task_v1 with groups: “AI apps can’t create backup tasks that use groups yet, because the sites in a group can change after you approve. Use the site IDs instead, or create the task in MainWP under Backups.”
  • create_backup_task_v1 with site addresses: “AI apps can’t create backup tasks from site addresses yet, because an address can point to a different site after you approve. Use the site IDs instead, or create the task in MainWP under Backups.”
  • update_backup_task_v1: “AI apps can’t edit backup tasks yet. Edit the task in MainWP under Backups.”
  • delete_backup_task_v1 with its backups, when there are too many to list: “This task has too many backups to list for approval, so an AI app can’t delete it together with its backups. Delete it in MainWP under Backups, or delete the task and keep its backups.”

Extension tools

MainWP reviews the tools of its released extensions and sets their tiers. A few examples:
  • replace_custom_dashboard_snippet_v1 from Custom Dashboard can install PHP code that runs on the Dashboard, so it’s always Dashboard, under every policy. The approval page shows the snippet’s language, size and SHA-256 instead of the code, so you can match it against the snippet you wrote.
  • Tools that carry a password or API key, such as the Matomo connection, are Dashboard. The Dashboard never stores the secret, so the action runs only when the app calls the tool again after your approval.
Extension tools MainWP hasn’t reviewed are listed in the panel under Not reviewed by MainWP. They’re always Dashboard tier (Confirm under Hands-off), and their approval page carries the heading “MainWP hasn’t reviewed which sites this changes”. Work they queue in the background isn’t cancelled when you revoke the connection. Each has its own switch, and the section has one too. Extension tools with missing or contradictory safety annotations are listed under Not available to AI apps and can’t be called. Tools waiting for an extension release appear under Not available yet. If you build an extension, Remote AI Connections for Extension Developers explains how your abilities get their tiers.

The three policies

You pick the policy on the consent screen, or when you create a connection token.

Standard

The default. Tools use the tiers in the table above. The consent screen describes it as “Risky changes need your confirmation in the app or your approval on this Dashboard.”

Strict

Every write and confirm tool moves to the Dashboard tier. Read tools still run directly. The consent screen describes it as “Every change needs your approval on this Dashboard.” Use Strict when you want to see every change on the Dashboard before it runs, for example on a connection from an app you’re still evaluating.

Hands-off

Dashboard-tier tools move to Confirm, so the app only needs a confirmation in the chat. These stay Dashboard under every policy:
  • add_site_v1
  • update_site_v1 when the call sets url, verify_certificate or ssl_version
  • replace_custom_dashboard_snippet_v1
  • restore_site_snapshot_v1, rollback_restore_v1 and cancel_restore_v1
  • delete_backup_snapshot_v1, cleanup_site_backups_v1 and delete_backup_task_v1
  • update_backup_settings_v1 when the call sets on_demand_retention_count
  • update_backup_destination_v1 when the call sets minimum_remote_retention_days
The first three could send a site’s keys to another address or run code on the Dashboard. The backup calls could overwrite a live site or remove the backups you’d recover from. Hands-off is offered only to users without Team Control limits. If the user gains limits later, the connection behaves as Standard until its policy is changed.
Hands-off gives up the one check that proves a person approved. The consent screen says: “Changes that would need your approval on this Dashboard only need a confirmation in the app. MainWP can no longer check that a person approved them, and instructions hidden in site content could lead the app to make them.” A post, comment or plugin description on a child site could contain such instructions.

Change a connection’s policy

In MainWP > AI > MainWP MCP > Connect directly, find the connection under Connections, pick a value in its Approval policy list and click Save policy.
  • Only the person who connected the app can change its policy or access. Other administrators can revoke it, nothing more.
  • Approvals waiting under the old policy stop working, and Dashboard-managed background jobs started under it are cancelled. The panel confirms with “The approval policy changed. Approvals waiting under the old policy no longer apply.”
  • A change to a looser policy (Strict to Standard, or anything to Hands-off) sends you the email “An AI app connection now needs fewer approvals”, followed by the app’s name. Administrators without Team Control limits also see a notice in the panel.

The approval page

  • Claude Code and VS Code ask you to open the link from inside the app.
  • Every other app gets the link in the tool result, with the text “This action needs approval in the MainWP Dashboard. Give the user this link”. The app should show you the link and call the tool again after you approve.
The link has this form, with the plain-permalinks form in parentheses:
Waiting approvals are also listed under Pending approvals in the panel, so you can open them from there.

What the page shows

The page, titled Approve an action from an AI app, shows:
  • the app and the date you connected it
  • the action
  • the sites it affects, and for an action on a client or tag, the sites it affects through them
  • the arguments the app sent; a password or key shows as “supplied”, never its value
  • the line “Only approve if you just asked your AI app to do this.”
You must be signed in as the user who owns the connection. A signed-out visitor is sent to the login page first, and anyone else is refused.

Approve or deny

Click Approve or Deny. Approve runs the action’s preview first, where it has one. For updates, the page then lists the exact items and expected versions. The page shows “You approved this. Go back to your AI app and ask it to try again.” Ask the app to call the tool again with the same arguments; the action runs once. Deny runs nothing. The app is told “The user denied this action.” once. If it calls the same tool again, a new approval starts. Some actions carry a password or key that the Dashboard doesn’t store. Their page shows “The Dashboard doesn’t store this action’s details, so its check runs when your AI app tries again.”

Time limits

  • An approval waits 30 minutes for your decision.
  • After you approve, the app has 15 minutes to call the tool again.
  • After that, the link reads “This approval can no longer be used. Ask your AI app to start again.”

When something changed

Before it runs, the Dashboard checks the action again. If a site was removed or you lost access to it, if the sites behind a client or tag changed, or, for tools that compare their preview, if the preview no longer matches what you approved, nothing runs. The page shows “The sites or changes are no longer what you approved, so nothing ran. Ask your AI app to start again.” The app’s next call creates a new approval. For updates, your approval covers the exact items you saw. An update that appears after you approved isn’t included. The child site installs the newest available version of each approved item, and the next sync shows what was installed.

Run it here instead

Some apps can’t call the tool again after you approve. After Approve, the page offers Run it here instead, which runs the approved action from the page, as you, without the app. Before it runs, the page checks the action again the same way the app’s retry would. If the preview changed, nothing runs, the approval can’t be used again, and the page shows “The sites or changes are no longer what you approved, so nothing ran. Ask your AI app to start again.” Otherwise the page shows the result, or “Started. The action runs in the background.” for a large action. If the app calls the tool again later, it gets “Already run from the Dashboard:” followed by the result. Run it here instead isn’t offered in these cases:
  • actions that carry a password or key, and extension tools MainWP hasn’t reviewed. The Dashboard keeps no runnable copy of these.
  • actions from a tool that compares its preview, when the approved plan is an update plan or has more than one step. The page can’t run that preview again, so the approval stays valid for the app’s retry.
In both cases the action runs only when the app calls the tool again.

Large actions

By default, an approved action over more than 10 sites runs as a background job, 10 sites at a time. The app gets a job ID and follows it with get_remote_job_status_v1. Before each step the Dashboard checks the connection again: revoking it, changing its policy, turning the activity log off or using the kill switch cancels the rest of the job.

The activity log

While Enable Network Activity logging is on, the Dashboard records every AI app call in the activity log, and changes an app makes show via MainWP MCP beside the user. While it’s off, the read tools MainWP has reviewed still run, but nothing records them. Changes and tools MainWP hasn’t reviewed are refused with “Changes and tools MainWP hasn’t reviewed need the MainWP activity log, so each call is recorded. Turn on Enable Network Activity logging under MainWP Settings > Network Activity Settings.”, and the approval page shows “Nothing ran: remote AI connections need the activity log turned on.”