Guides

Permission modes and approvals

Mentat has five permission modes, and each one sets the tool's own permission setting; Mentat adds no approval layer of its own. Shift + Tab in the composer cycles them, and Approve edits is the default.

Checked against Mentat's code on .

The five modes

ModeWhat it does
Approve editsApplies file edits automatically and asks before untrusted commands.
Approve for meHands decisions to the tool's own automatic reviewer.
Ask every timeAsks before any write or untrusted command.
Plan firstThe agent researches and presents a plan before it builds anything.
Approve everythingRemoves the sandbox and the approval step. Use it only where the agent can do no harm.

The mode for new turns is set under Settings, Agents, Default permissions. Plan first is skipped for a tool that does not offer it.

What each mode sets in each tool

ModeClaude CodeCodexGrok Build
Approve editsacceptEditsAsks unless trusted; may write in the workspaceacceptEdits
Approve for meautoCodex's automatic reviewer; may write in the workspaceauto
Ask every timeClaude Code's asking modeAsks unless trusted; read-onlydefault
Plan firstplanAsks unless trusted; read-onlyplan
Approve everythingbypassPermissionsNever asks; full accessbypassPermissions

Approval cards

When an agent needs permission, the card shows the request in full: the folder, the command or the file changes, any blocked path, and the tool's raw request. It is headed by what is being asked: Allow command?, Allow file changes?, Allow requested permissions?, Allow MCP tool? or Open requested link?

The choices are the ones that tool offers, in its words.

  • Claude Code: Allow, Deny, Deny and stop, and, when Claude proposes a rule, Allow and remember for this session, this project or all projects.
  • Codex: only the decisions Codex lists for that request, such as Allow once, Allow this command for this session, or Always allow commands starting with a given prefix, beside Deny and Deny and stop.
  • Grok Build: the agent's own options, such as Allow once, Always allow and Reject.

Questions and plans

A question card lists the agent's options and a field to describe your own answer. A plan comes back as a card to review before anything is built. For Claude Code the choices are Approve and edit automatically, Approve, ask before edits, and Keep planning, with room for feedback; for Codex and Grok Build they are Implement and Keep planning.

Sub-agents

A request from a sub-agent surfaces in the composer of the chat that started it, marked Requested by with the sub-agent's name, so you answer it where you are already working. The oldest pending request is shown first.

For engineers who stay close to the code

Set your agents in motion

Free to use, with no card. Every new account starts with fourteen days of Pro.

macOS 13 or later, Apple silicon. Windows is next. Git 2.45 or later. At least one of Claude Code, Codex or Grok Build, installed and signed in.