> ## Documentation Index
> Fetch the complete documentation index at: https://vespin.dev/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Permissions

> Choose what an agent may do without asking, and answer its approvals

An agent runs in one of three permission modes. In the two that ask, it stops before an action and waits in its transcript for you to allow it.

## Modes

| Mode | Claude Code asks before | Codex asks before |
| - | - | - |
| **Ask for approval** | Every file edit and shell command, bar [always allowed commands](#always-allowed-commands) | Writing outside its checkout or using the network |
| **Approve for me** | Only actions Claude Code flags as unsafe | Nothing. Codex's own review answers for you. |
| **Full access** | Nothing | Nothing |

New agents start on **Full access**. Change your default in **Settings › Agents › Default permissions**. To change one agent, open its model chip's menu and choose **Permissions**, before it starts or while it runs. A change takes effect from the next turn. An agent in an asking mode shows the mode's icon on its model chip.

An automation's agents always run on Full access.

## Approvals

A waiting action stands in the agent's transcript with the command or edit it would make.

![An agent's transcript: Run the members tests waits for approval over the command pnpm test src/members, with Allow, Always allow pnpm test \*, and Deny under it, and Waiting on you · 1 approval below](https://mintcdn.com/stage-a8145d8f/EAMsVV_TcNHcjtNJ/images/permissions-approval.png?fit=max\&auto=format\&n=EAMsVV_TcNHcjtNJ\&q=85\&s=e829abcc9562e1e47ce57014158caca0)

* **Allow** runs it once.
* **Always allow** runs every action like it without asking for the rest of the agent's session. When Claude Code proposes a command prefix, such as `pnpm test *`, the button names it, and the prefix is kept for all your Claude Code agents.
* **Deny** refuses it, with an optional note telling the agent why.

Sending the agent a message denies the waiting action, and **Stop** withdraws it. A request has no timeout. It expires only when its turn ends.

## Always allowed commands

**Settings › Agents › Always allowed** lists the command prefixes your Claude Code agents run without asking, in any mode. The list is yours alone, holds up to 200 commands of up to 8 words each, and reaches each agent the next time it starts up. Add one with **Add command**, or with **Always allow** on an approval.
