Permission modes
Choose how much an agent can do on its own in each chat: Approve, Autopilot, Plan, or Bypass permissions — and which checks no mode skips.
A permission mode decides how much an agent can do without asking you first. Each chat keeps its own mode, so you can let one chat run a long literature sweep on its own while another asks before every edit to your thesis. For the rules and trust settings that work alongside modes, see permissions.
Available modes
| Mode | What runs without asking | Best for |
|---|---|---|
| Approve | Very little — you approve the agent’s actions as it goes | Important chapters, unfamiliar tasks, learning how agents work |
| Autopilot | Routine work. Shell commands, installs, deletions, protected files, and anything outside the project still ask | Everyday writing and research |
| Plan | Reading only. Nothing is edited until you approve the plan | Big changes you want to agree on first |
| Bypass permissions | Almost everything, without stopping to ask | Work you fully trust, in a project you can restore |
Checks no mode skips
- A Deny rule blocks its action in every mode.
- Commands that delete files always ask, and can’t be set to “always allow”.
- In plan mode, any command that isn’t read-only asks first — even with Bypass permissions on.
- A project you haven’t trusted stays in Restricted Mode, whatever the chat’s mode. See trust a project folder.
Common setups
| You want to… | Use |
|---|---|
| See and approve every action on an important chapter | Approve |
| Get on with everyday writing without constant prompts | Autopilot |
| Agree on a restructure before any text changes | Plan, then approve the plan |
| Run a long data-processing job without approving every command | Autopilot with /sandbox on macOS — see permissions |
| Let a long literature sweep run while you edit elsewhere | Autopilot in that chat, Approve in your editing chat |
Which mode a chat starts in
- New chats start in the mode set in Settings → AI Agent → Permission mode.
- New installs start in Autopilot. If you used Multilo before modes moved to each chat, your mode didn’t change.
- After that, each chat keeps its own mode — including when you reopen it later.
Switch a chat’s mode
- Plan mode: press Shift+Tab in the chat box, type
/plan, or choose Mode → Plan in the chat’s settings menu (the sliders icon). - The default for new chats: Settings → AI Agent → Permission mode.
A setting you change mid-chat applies from your next message.
Approve mode
The agent stops for your approval as it works, so you see each action before it happens. It’s the most careful mode: slower, but nothing surprises you. A good choice for a chapter your supervisor has already approved, or the first few times you use an agent on a task.
Autopilot
Routine work runs straight through — reading, drafting, editing your documents in the project. The agent still asks before shell commands, installs, deletions, protected files, and anything outside the project. This is the everyday mode: you review the result with the review bar instead of approving each step.
Plan mode
The agent reads and proposes a plan, but edits nothing until you click Approve & run. Read-only commands run normally; anything else asks first. Plan mode stays on until you approve, so you can ask for changes to the plan as many times as you need. See planning.
Bypass permissions
The agent doesn’t stop to ask. Turning it on asks you to confirm. Even in Bypass, Deny rules still block, deleting files still asks, and plan mode still asks before commands that aren’t read-only.
Use Bypass with care
Only use Bypass permissions for work you fully trust, in a project that’s inside Multilo’s history and backed up with Git, so any change can be undone. Never use it on a project you just downloaded from someone else.