# Permissions

> Control what the KazoCode agent can do without asking. Choose a permission mode, save allow and deny rules, approve folders, and see which actions always ask.

Every tool the agent uses has a risk level: **read**, **write**, **execute** (commands) or **network**. Reading inside your project never asks. For everything else, your permission mode and saved rules decide whether KazoCode asks you first.

## Permission modes

Choose a mode from the menu under the message box.

| Mode | What it does |
| --- | --- |
| **Default permissions** | Uses your saved rules and asks for anything they don't cover |
| **Assisted permissions** | The current model reviews each tool call. Safe, reversible steps inside the project run without asking; anything else still asks you |
| **Accept edits** | File changes in the project are approved automatically and stay undoable. Commands, web access and other actions still ask |
| **Allow all** | Approves every tool call and retries on errors |
| **Autopilot** | Approves every tool call and keeps going until the task is done. This can use more of your model budget |

> [!WARNING]
> **Allow all** and **Autopilot** run commands without asking. Use them in projects you trust, and keep the work in Git so you can review it afterwards.

## Approving a request

When the agent asks, the prompt says exactly what it wants to do, such as the command it will run, and shows the diff for a file change. Choose **Allow** to let it run this once, or **Decline** (`Escape`) to say no. When KazoCode can tell what similar requests look like, a third button allows this one and stops asking for matching requests from now on, by saving a rule.

After declining, send a message to tell the agent what to do instead.

## Saved rules

Rules you save live in the project's `.kazo/permissions.json` (shared with anyone who uses the project) or in your user settings (just for you). A deny rule always wins over an allow rule. Use `/permissions` to list, remove or clear them.

## Actions that always ask

Some actions ask no matter which mode you're in:

- Deleting files and overwriting documents.
- Destructive commands, such as `rm -rf`, `git reset --hard`, `git push --force`, `git clean`, `git branch -D` and their equivalents on Windows.
- Working outside the project folder. Approve a folder once to let the agent work there.

Deleted files go to the Recycle Bin or Trash when your system has one.

## Sandboxed commands

`/sandbox` runs the agent's shell commands confined to the workspace, with or without network access, so even approved commands can't reach the rest of your computer.

## Audit log

Every tool call the agent makes is written to an audit log, so you can see afterwards exactly what ran and what was approved.
