Skip to main content
Actions fire after the Agent decides. Unlike Tools (which the Agent picks during a run), Actions run based on wiring: the Flow decides whether to fire, the Agent’s output decides what to pass in. Same command shape as Tools, minus parameters (Actions don’t have LLM-picked args).

Commands

actions create

Flags:
  • --name <n> (required).
  • --description <d>.
  • --runtime <json> — HTTP or inline runtime spec (same shape as Tools).
  • --x <n>, --y <n> — canvas position.

actions update

Update flags

Identity
  • --set-name <n>
  • --set-description <d>
  • --set-icon <name>
Approval
  • --set-approval on|off — require approval before firing.
HTTP runtime
  • --set-runtime-kind http|inline
  • --set-http-method <m>
  • --set-http-url <url>
  • --set-http-auth-kind <k>
  • --set-http-auth-header <name>
  • --set-http-auth-value <v>
Inline runtime
  • --set-inline-lang <lang>
  • --set-inline-code <text> / --set-inline-code-file <path>
Position/size
  • --set-x <n>, --set-y <n>, --set-width <n>
Bulk
  • --patch <json>

actions edit

actions delete

Removes the Action; wires prune.

actions headers

Same shape as Tools:
Secret values via secrets://<ref>. Interactive with --value ?. See Piping & secrets.

Tool vs. Action — quick refresher

Rule of thumb: if the Agent might need the result to reason further, it’s a Tool. If it’s fire-and-forget after the answer, it’s an Action.

Recipes

Add a Slack notification with an approval gate

Every Slack post now waits for approval.

Migrate a header rotation