Skip to main content
This starts an agent in its own process, in the current folder. It keeps going with the app, the terminal and the island closed. It prints the run’s id.
LimitsLinux and macOS only. On Linux it needs bwrap. Headless runs use Workers AI models for now.

The sandbox

Its tools run inside a sandbox: bubblewrap on Linux, Seatbelt (sandbox-exec) on macOS. The workspace is writable and the rest of the machine is read-only. There are no credentials, and no network unless you allow it. --allow grants more, as a comma-separated list: --read <path> lets the sandbox see one more path, read-only. Repeat it for more paths. The model’s API calls run outside the sandbox, in the agent’s own process. So does what the agent can’t do from inside, such as sending mail or pushing: Arakun does it for the agent, and only after you approve it.
--no-sandbox runs the tools on your machine, without a sandbox. Arakun waits for you to confirm it first; --yes confirms it at once.

The policy

--policy decides which tool calls need your approval: The policy only saves you from answering for harmless calls. The sandbox is what keeps the agent in. When the policy can’t tell what a command does, it asks. A call that asks waits for your answer:
A call nobody answers within 24 hours is denied.

Follow and manage runs

The runs also show in the sidebar, under the project that holds their folder. One opens read-only. A message you send to it goes to its own process, which starts again if it had finished, so it stays in its sandbox.

More options

With --wait, nobody is there to answer, so calls that would ask are declined. --wait exits with 0 when the turn ended on its own, 1 on an error, 124 on the timeout and 130 on a signal. --wait --no-sandbox needs --yes. A headless run uses memory like any other session. Its sandbox only sees the project’s skills, not those in your home folder.