# Permission modes

Every agent can run shell commands and edit files in its session's folder. The permission mode
decides what it asks you about first. When it asks, the question shows in the transcript with the
tool's name and exactly what will run: the command, or the file it changes. You answer **Allow** or
**Deny**. A Deny stops the turn. Once answered, the question shows **Allowed** or **Denied**.

## The five modes

| Mode                   | What happens                                                     |
| ---------------------- | ---------------------------------------------------------------- |
| **Manual**             | The agent asks before each edit and command.                     |
| **Edit automatically** | File edits go through; commands still ask.                       |
| **Plan**               | The agent explores and plans, and edits nothing.                 |
| **Auto**               | Safe steps go through; anything risky is checked before it runs. |
| **Bypass**             | Nothing asks; the agent edits and runs commands on its own.      |

A new session starts in Manual. A fork, `/branch` or `/subtask` starts in the mode of the session
it came from.

## Which agent has which

|             | Manual | Edit automatically | Plan | Auto | Bypass | Switching               |
| ----------- | ------ | ------------------ | ---- | ---- | ------ | ----------------------- |
| Claude Code | yes    | yes                | yes  | yes  | yes    | any time, even mid-turn |
| Codex       | yes    | no                 | yes  | yes  | yes    | between turns           |
| Gemini CLI  | yes    | yes                | yes  | no   | yes    | any time                |

Codex has no Edit automatically, because none of its settings lets edits through while still asking
for every command. Gemini CLI has no Auto. With Codex, Manual still runs the commands Codex knows
only read, such as `ls` and `cat`, without asking.

Auto is each agent's own. With Claude Code, it's Claude Code's auto mode, and Claude Code decides
what goes through. With Codex, the approvals Codex would ask you for go to its own review instead.
A command the review denies shows in the transcript as a note, with the reason where Codex gives
one, and **Approve once** lets the agent try it again on its next turn, this one time only.
`/approve` lists the actions the review denied in this session, newest first, each with **Approve
once**. See [Conversation commands](/docs/conversation-commands/).

Your account can still refuse a mode the agent normally has, such as Claude's Auto. Then the
switch fails with the agent's own reason, and the session keeps its mode.

## Switching

- The mode picker is in the composer's toolbar. Modes the agent doesn't have are shown, but
  disabled, and the current one has a check.
- **Shift+Tab** in the message box steps through the modes the agent has.
- Switching to Bypass always asks you to confirm first: "Turn off permission prompts?", with
  **Switch to Bypass** and **Cancel**.
- When a switch can't happen now, the picker is disabled and says why beside it. While a Codex turn
  is running, it says "The mode can change once this turn ends."
- Each change shows as a quiet row in the transcript.

The agent can also change its own mode. When you approve a plan, Claude leaves Plan by itself, and
the picker follows.

## Questions, plans and other asks

Besides asking before a tool, an agent can ask you a question, show you a plan, ask you to fill in a
form, or ask you to open a link. No permission mode answers these: they wait for you, even in
Bypass. How each one shows, and how you answer it, is in
[Questions and plans](/docs/questions-and-plans/).

|             | Questions         | Plans | Forms | Links |
| ----------- | ----------------- | ----- | ----- | ----- |
| Claude Code | yes               | yes   | yes   | yes   |
| Codex       | in Plan mode only | yes   | yes   | yes   |
| Gemini CLI  | —                 | yes   | —     | —     |

Forms and links come from MCP servers. A form with a field Moontors can't show is declined.

### Questions

- **Claude Code** asks one to four questions at once. A question can take several choices, and its
  choices can come with a preview. **Other**, for an answer of your own, is always there.
- **Codex** asks for one choice per question, with **Other** where Codex offers it.

With either agent you can add a note to your answer.

### Plans

**Accept this plan?** offers an answer for each mode the agent can go on in, such as **Yes, and
edit automatically**. **Yes, and bypass permissions** confirms first, as switching to Bypass does:
"Turn off permission prompts?"

- **Claude Code** goes on in any of its modes but Plan.
- **Codex** goes on in Manual or Auto, with its own prompt to carry out the plan.
- **Gemini CLI** goes on in Manual only. Moontors reads its plan from Gemini CLI's plan file, and
  your feedback goes to it as your next prompt.

**No, keep planning** sends your feedback instead, and the agent stays in Plan.

## Allow for this session

Where the agent supports it, a question also offers **Allow for this session**. Its tooltip names
the exact rule it adds. Answering with it adds that rule, so the same request isn't asked again in
this session, and the question shows **Allowed for this session**. The session's rules are listed
in the Configuration dialog, under This session, where you can remove one. See
[Configuration](/docs/configuration/).