Automations
Create one
Open More at the bottom of the sidebar, then Automations, and click New automation.
Name it and pick what it runs: AI prompt, Command or Action. For a prompt, choose the agent, model and effort at the right of the box.
Pick where it runs. Standalone runs in your home folder; a project runs in its folder, in a fresh copy, or in a new Git worktree that leaves your real folder untouched.
Choose Full access or Read only. Nobody is there to approve its steps, so pick Read only for a job that should only look.
Set when it runs, then click Create automation. Run now starts it straight away.
Read a run and reply
Every run keeps its outcome and how long it took. Open one to read the agent’s answer as a conversation, and reply there to keep it going. A running job shows its output live.
Let it ask you first
In the prompt, say when the agent should ask, for example before it picks a version number. There is nothing to switch on. When it needs you, the Automations row says Needs you and a notice shows the question; one click answers it, and the agent carries on in the same conversation. Until you answer, the automation waits, and a run that comes due is skipped.
Good to know
- Automations fire only while lpm is running on an awake computer. For jobs overnight, turn on Keep awake or run them on a computer that stays on.
- Only one run of a job goes at a time. A run that comes due during it is skipped.
- Add a check, such as “are there new commits?”, and the job runs only when it passes. A verify command, such as your tests, marks each run passed or failed.
- Full access is the default. OpenCode jobs always run with full access.
- Each automation can use the Claude account you pick for it.
- Agents and scripts can manage them too, with
lpm automations, and lpm Link can create, run, pause and answer them from your phone.
Read more: scheduling Claude Code tasks.