Permissions and Safety
After reading this page you will be able to: pick the right permission mode, understand every approval prompt, and stop the AI at any time.
1. Permission modes: how much power the AI gets
Section titled “1. Permission modes: how much power the AI gets”Every action the AI takes (editing files, running commands, calling tools) is governed by the current permission mode.
Where: the shield-shaped Permission pill at the lower right of the chat input box (the home page input area has one too)
The available modes are reported by each assistant — the dropdown always shows what your current assistant supports:
| Mode | Meaning | Best for | Risk |
|---|---|---|---|
| Default | Every tool call needs your confirmation | Beginners, sensitive projects | Almost none — just a few extra clicks |
| Auto Edit | File edits and info tools are approved automatically; commands still ask | Trusting AI with code, fewer clicks | Review what was changed |
| YOLO | Everything is approved automatically, no questions | Advanced users, isolated environments | High — review the plan first |
Other assistants may offer different modes, commonly: Plan Mode (produces a plan first and acts only after you approve), Read Only (looks but never touches), Accept Edits, Full Access, and Full Auto (No Sandbox). Hover a menu item to see its description when unsure.
2. Switching the permission mode
Section titled “2. Switching the permission mode”- Click the shield Permission pill to open the Permission Mode menu
- Click the mode you want; a ✓ marks the active one
- “Mode switched successfully” means it is in effect; some assistants say “Takes effect on the next turn” — the menu shows an ⏱ marker on that mode, and the AI switches over automatically once it finishes the current turn
3. Approval dialogs: every grant is visible
Section titled “3. Approval dialogs: every grant is visible”In Default mode, the chat shows a confirmation card whenever the AI is about to act:
| Prompt | When it appears |
|---|---|
| Apply this change? | Editing a file (the diff is shown right in the card) |
| Allow execution? | Running a command |
| Do you want to proceed? | Other actions that need confirmation |
| Allow execution of MCP tool “Y” from server “X”? | Calling an MCP tool |
| Button | Effect |
|---|---|
| Yes, allow once | Grant this one time only |
| Yes, allow always | Similar actions stop asking one by one |
| Yes, always allow tool “Y” from server “X” | Grant that single tool of that server |
| Yes, always allow all tools from server “X” | Grant every tool of that server |
| No (esc) | Decline this request; the AI will try another way or stop |
4. Thinking level
Section titled “4. Thinking level”Assistants that support it show a Thinking Level dropdown next to the Permission pill; it controls how deeply the model reasons before answering. Switching is confirmed with “Thinking level switched”. When unsure, keep the default.
5. Stopping a task
Section titled “5. Stopping a task”- While the AI is working, the send button becomes a Stop button (square icon) — click it to halt the current task immediately
- Desktop Automation tasks are stopped together with the external automation, so the chat stops and the mouse never keeps moving on its own (see Automation Assistant)
6. Safe habits
Section titled “6. Safe habits”- For sensitive jobs (accounts, payments, deleting files), use Read Only or Plan Mode and let the AI advise only
- Before switching to YOLO, read the plan it produced, then let go
- You can always click Stop, or reject a single action by closing the confirmation card
What success looks like
Section titled “What success looks like”- Clicking the Permission pill opens the Permission Mode menu and a switch ends with “Mode switched successfully”
- You see a confirmation card before the AI acts and can grant with “Yes, allow once” or decline with “No (esc)”
- The Stop button appears while the AI works and clicking it halts the task immediately