> ## Documentation Index
> Fetch the complete documentation index at: https://wonder.design/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# Permission modes and approvals
> Choose how much your agent does without asking, answer its approval requests, and review the rules it follows without asking.
A chat's mode sets how much its agent does on its own before it stops to ask you. Pick a mode for each chat, choose the one new chats start in, and answer the agent when it asks.
**Before you start** · Desktop app · Free plan
## The four modes
| Mode | Claude Code | Codex |
| - | - | - |
| **Ask** | Asks before edits and commands. | Works in this folder; asks for broader access. |
| **Edit automatically** | Edits freely; asks before commands. | Not offered. A Codex chat set to start in this mode starts in **Ask**. |
| **Auto** | Asks only about risky actions. The agent's own reviewer decides what's risky. | Asks only about risky actions. The agent's own reviewer decides what's risky. |
| **Full access** | Edits and runs without asking. | Edits and runs without asking. |
In every mode, the agent's questions to you still wait for your answer. **Auto** isn't available for every model; the menu says **Not available for this chat's model** when it isn't.
## Choose a chat's mode
1. In the chat's message box, click the mode next to the model. It shows the chat's mode, such as **Auto**, or only its icon when the chat is narrow. Hover it to see the mode's name.
2. Pick a mode.
If the agent is working, the menu says **Applies from your next message.** Changing a chat's mode also sets the mode new chats in that repository start in. **Full access** is the exception: after you pick it, the repository's new chats start in your **Default**.
## Choose the mode new chats start in
Open **Settings > Agents > Mode for new chats**.
* **Default** is the mode new chats start in. It's **Auto** until you change it.
* Each repository has its own row. **Follow the default** keeps it on **Default**. Pick **Ask**, **Edit automatically** or **Auto** to give the repository a mode of its own.
A repository can't keep **Full access**: its new chats start in **Full access** only when it follows a **Default** of **Full access**.
## Approve or deny an action
When the agent needs your go-ahead, the chat shows a card such as **Run this command?**, **Change this file?** or **Reach this site?**.
* Click **Allow**, or press ⌘ ↵. When there are other ways to allow, the main button says **Allow once**, and the menu next to it offers the rest:
* A rule so the agent stops asking, for example **Don't ask again for "npm test" commands in this repository**. Some rules last only for this chat, such as **Allow for this chat**.
* **Allow and switch to Auto** or **Allow and switch to Full access**, which also changes the chat's mode.
* Click **Deny**, or press Esc. The menu next to it offers **Deny with a note**, to tell the agent what to do instead, and for a command, **Edit the command instead**, to change it before it runs.
## Review what's allowed without asking
**Settings > Agents > Allowed without asking** lists the rules saved when you chose "don't ask again", in Wonder, Claude Code or Codex.
* Claude Code rules are listed for each repository, and Codex rules under **Every repository**.
* Click **Open in editor** to see the file a rule is saved in.
* To remove a rule, click **Remove** next to it, then **Remove rule**. The agent asks again before doing what the rule allowed, in Wonder and outside it.
## Let agents create files, folders and pages
**Let agents create files, folders and pages**, in **Settings > Agents > Access**, is on by default: agents create the Wonder files, folders and pages a task or a Preview needs, instead of asking. Turn it off and agents can't create them; they ask you to. The change applies when a chat's agent next starts.
## Chats that worked on a fork's pull request
Anyone on GitHub can write a fork's code, so a chat that worked on a pull request from a fork, or took in text from such a chat, runs contained to keep your computer and settings safe. The chat says **This chat now runs contained** and shows a **Contained** chip in its message box.
A contained chat:
* Can't use **Full access**, and asks before acting.
* Changes files only in its own folders, reaches only listed sites and asks you before anything else. Where Wonder can't sandbox the agent, such as on Windows or when your own agent settings loosen the sandbox, it asks before each command and before changing a file outside its folders instead.
* Asks every time and never remembers your answer, so its approval cards offer no "don't ask again" choices.
In a checkout of a fork's pull request, Claude chats also load only your own Claude Code settings, not the repository's, so its MCP servers and hooks don't run.
Click **Contained** to see **What it was stopped from doing**. A chat stays contained; for work that doesn't involve the fork, start a new chat.
## If something goes wrong
* **The chat asks even though its mode shouldn't.** The **Mode** menu says why, for example **The agent's own settings don't allow Full access, so this chat asks before acting.** or **The agent isn't running Auto in this chat, so it asks before acting.** Change the agent's own settings or pick a model that offers **Auto**.
* **A mode is greyed out.** **Not in a chat that has worked in a fork.** means the chat is contained. **Not while this chat's worktree or repository is removed.** means the chat's folder is gone; it can only use **Ask**.
* **"Couldn't change the chat's mode. Try again."** Pick the mode again.
* **"Wonder can't read this file as rules."** The rules file has an error. Click **Open in editor** to fix it.
## Related
* [Choose models and reviewers](https://wonder.design/docs/choose-models-and-reviewers.md)
* [Choose what a chat works with](https://wonder.design/docs/choose-what-a-chat-works-with.md)
* [Agent and chat preferences](https://wonder.design/docs/agent-and-chat-preferences.md)
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.