Permissions
The Integration Kit asks before it does anything destructive. The permission model is simple and mirrors how Claude Code’s own CLI permissions work, so if you’re already a Claude Code user, it’ll feel identical.
Which tools ask for permission
| Category | Asks? |
|---|---|
Read-only (read_asset, read_blueprint, find_assets, list_*) | No |
Edit (edit_blueprint, edit_material, set_pin_value, …) | Yes, but Allow for session is usually what you want |
Create (create_blueprint, create_material_instance, …) | Yes |
| Delete / Rename / Move | Always, every time |
File operations (write_file, delete_file, rename_file) | Always |
| PIE / Runtime control | Yes, once per PIE session |
The permission card
When the AI tries a destructive tool, the chat shows a card with:
- Tool name (for example,
delete_asset). - Target (for example,
/Game/Deprecated/OldTex.uasset). - Inputs, the parameters the AI wants to pass.
- Three buttons: Allow, Allow for session, Deny.
Allow
Approves this one call. The next similar call will ask again.
Allow for session
Approves this call and all similar calls in the current chat session (until you close the editor or switch tools). Use this for routine edits. You approve once and the agent runs uninterrupted.
Deny
Blocks this call. The AI sees the denial, typically explains what it was trying to do, and either asks for a different approach or reports back without that step.
Session scope
“Allow for session” scopes to:
- This chat.
- This active tool (switching to a different integration resets the approval cache).
- This editor process (closing UE clears the cache).
If you want to reset approvals without closing the editor, start a fresh chat (Ctrl+N). The new chat starts with a clean approval list.
Safety by default
The defaults are conservative:
- The Integration Kit ships with every category enabled, but destructive sub-tools still prompt.
- File Operations is enabled but
write_file/delete_file/rename_fileall prompt. - PIE control is enabled but asks once per PIE session.
You’d have to actively click Allow for session or Bypass mode to let the AI run destructive tools without confirmation.
Review the diffs
The chat thread is an audit trail. Every tool call is visible with inputs and outputs. If a big refactor just ran, scroll through the tool cards before moving on. Unreal’s own Undo stack also works for most asset edits.
Typical workflows
- Exploratory chat. Default permissions; approve each destructive step. Slower but lets you intercept.
- Known refactor. Approve the first destructive step “for session”. Rest of the refactor runs hands-off.
- Review-only chat. Switch to Plan Mode; the AI describes the change without running anything. Useful for debugging AI-driven regressions.
- Hands-off overnight run. Switch to Bypass (if your activated tool supports it). The AI runs everything without asking. Review the tool cards when you come back.