# Sandboxing

URL: https://docs.hanzo.ai/docs/dev/configuration/sandbox

> How Hanzo Dev sandboxes command execution for safety.

## Sandbox & Approvals

Hanzo Dev enforces sandboxing to prevent unintended side effects when the agent executes commands.

### Sandbox Modes

| Mode                 | Description                                            |
| -------------------- | ------------------------------------------------------ |
| `off`                | No sandboxing (commands run directly)                  |
| `workspace-read`     | Read-only access to workspace                          |
| `workspace-write`    | Read/write access to workspace (default for full-auto) |
| `network-off`        | Disable network access                                 |
| `danger-full-access` | Full access (use with caution)                         |

### Approval Policies

| Policy       | Description                      |
| ------------ | -------------------------------- |
| `always`     | Always ask for approval          |
| `on-failure` | Ask only on command failure      |
| `on-request` | Ask when model requests approval |
| `never`      | Never ask for approval           |

### Usage

```bash
# Read-only mode (safest)
dev --sandbox workspace-read "analyze this codebase"

# Write mode with approval on failure
dev --sandbox workspace-write --ask-for-approval on-failure "refactor"

# Full auto (sandboxed writes, auto-approve)
dev --full-auto "fix all linting errors"

# Bypass all safety (DANGEROUS)
dev --dangerously-bypass-approvals-and-sandbox "your prompt"
```

### Mid-Session Changes

Change approval level during a conversation with `/approvals` in the TUI.