Skip to main content
Select the computer a conversation may use before asking the agent to run work there. It can use that computer’s existing CLIs, files and sign-ins, so name the intended device and directory when they matter. An online device is not proof that it is the right device or cloud account.

​
A command that finishes

For inspecting files, running git, installing dependencies, or running a build, the agent can use local_exec. It returns output and an exit status. Commands that run too long may be stopped; use a terminal for processes meant to stay open.

​
A shared terminal

Ask for a terminal when you want to see and interact with a shell, development server, watcher, log tail or CLI waiting for input. local_terminal opens it in the dock. You and the agent can type into it. Stop the foreground process or close the terminal when done; closing the tab ends its shell. Terminal output is a recent snapshot, not a completed command’s exit status. A running terminal is not a durable schedule or monitor. Use triggers for future work that must be registered persistently.

​
Approval and access

Local actions follow the conversation’s permission mode. In Auto Mode, the runtime decides which actions need permission. Full Access approves available permission requests automatically; local_exec has no extra universal “ask before every command” gate. Local commands run with the computer user’s permissions, including existing cloud and repository sign-ins. Verify that identity before cloud administration. The credential picker for connected tools does not remove logins already on a local machine. See Runtimes and computers before sharing one.