Skip to content

Permissions

The real permission-mode identifiers for each agent, and what they mean.

Each agent defines its own permission modes, controlling how much an agent can do without asking you first. AgentDock surfaces these as-is — it doesn't invent a unified permission model on top of them.

Claude Code

ModeMeaning
defaultStandard behavior — prompts for permission per the CLI's own rules.
acceptEditsFile edits are accepted automatically; other actions may still prompt.
planThe agent can propose changes but not write them.
bypassPermissionsNo permission prompts. Use with caution.
dontAskSuppresses prompts for actions that would otherwise ask.
autoAutomatic mode as defined by the CLI.

OpenAI Codex

ModeMeaning
defaultStandard sandboxed behavior.
read-onlyThe agent can read but not write to the workspace.
workspace-writeThe agent can write within the workspace sandbox.
danger-full-accessNo sandboxing. Use with caution.
bypassNo permission prompts. Use with caution.

Codex's mode is set when a thread is created and does not switch live mid-session, unlike Claude Code's.

Google Antigravity

ModeFlagMeaning
default(none)Standard behavior.
accept-edits--mode accept-editsFile edits are accepted automatically.
plan--mode planThe agent can propose changes but not write them.
bypass--dangerously-skip-permissionsNo permission prompts. Use with caution.

Warning

Modes named bypass, bypassPermissions, or danger-full-access remove the safety prompts entirely. Only use them in a workspace you trust, ideally under version control so you can review or revert changes.

How permission requests appear

When an agent hits an action it isn't already permitted to take, AgentDock shows it inline in the conversation as an interaction card and waits for your response — it does not use a separate global approval dialog. For the broader safety model this fits into, see Permissions & Safety: agent permissions.