Interaction modes
Interaction modes control how much autonomy an agent has when executing tools. You can switch modes at any time during a conversation using the toggle at the bottom of the chat panel.
The three modes
| Mode | Behavior | Best for |
|---|---|---|
| Normal | Agent asks before each tool execution. You approve or deny. | Exploratory work, unfamiliar codebases, learning what the agent does. |
| Plan | Agent proposes a plan before executing. You approve or reject the plan. | Structured tasks where you want to review the approach first. |
| Auto-Accept | Tools execute immediately without confirmation. | Trusted workflows, repetitive tasks, batch operations. |
Switching modes
The interaction mode toggle is a segmented control at the bottom of the chat panel. Click a mode to switch, or press Shift+Tab to cycle through available modes.
The active mode is highlighted:
- Normal: orange
- Plan: blue
- Auto-Accept: purple
A brief ring pulse appears when switching to Plan or Auto-Accept as a visual confirmation.
Availability
All three interaction modes are available on every tier, including Starter (free). These are standard Claude CLI permission modes and are not gated.
Start with Normal mode to understand what tools an agent uses for a given task. Once you are comfortable with the pattern, switch to Auto-Accept to speed things up.
Interaction mode vs. permission mode
These are separate controls:
- Interaction mode (Normal / Plan / Auto-Accept) controls the approval flow, i.e. whether the agent asks before each tool use
- Permission mode (Normal / Bypass) controls the safety boundary, i.e. whether the CLI enforces permission checks at all
You can combine them independently. For example, an agent in Normal interaction mode with Bypass permissions will still prompt you in the UI, but the underlying CLI will not add its own permission checks.
Auto-Accept + Bypass Permissions gives the agent full autonomy with no safety checks. Only use this combination when you fully trust the agent's task and context.