DOCUMENTATION
Docs › Permission levels & safety
Permission levels & safety
You decide how far Claude may go. Anything that changes the live site waits for your Approve.
The three levels
| Level | Claude can | Needs your approval |
|---|---|---|
| Read-only (default) | Read pages, posts, plugins, mind maps and site health | — |
| Content editor | Create and edit drafts, upload images, update mind maps | Editing live pages, publishing, menus, homepage and site title |
| Site maintainer | Everything above, plus theme settings, widgets and plugin installs | Plugin installs and updates (blocked while the lab lock is on) |
For a whole website, one build plan replaces the separate approvals for everything it lists.
Seven safety rules built in
- Its own user. Claude acts as
claude-agent(Editor), never as an administrator. - One door. Every action passes a single gateway that checks the level, the locks and the per-tool switches.
- Approval queue. Live changes are prepared, shown with a diff, and only run when an administrator clicks Approve.
- Undo. A revision is saved before every edit; mind maps keep their last five versions.
- Activity log. Every call is written down with its result and an undo link.
- Locks. The plugin obeys AI Engine’s emergency write lock and Online Lab lock unless an administrator turns that off.
- Kill switch. One click disconnects Claude and revokes every sign-in.
