SCHEDULED AI WORKFLOWS

Schedule agent work
on your own computer.

Caterm workflows run ordered prompts through Codex or Claude Code on a connected device. Start a run manually, schedule it once, repeat it at a chosen interval or calendar time, or loop after each successful completion.

Create your first workflow

1. Connect and configure an agent on your Mac or Linux computer.
2. Open Workflows → New workflow. Choose a name, device, agent, and relative project folder.
3. Add between 1 and 50 named prompt steps. They run in order using the same agent session within each run.
4. Review the execution permissions, repeat mode, device waiting window, and timeout per run before saving.
5. Choose Run now or let the schedule start the workflow. Open Workflow runs to follow progress, answer questions, handle approval requests, or stop a run.

Supported schedules

Choose manual only, a one-time run, a fixed interval, daily, weekdays (Monday–Friday), weekly, or an advanced five-field cron expression with minute precision. Weekdays skip Saturday and Sunday in the selected time zone, without holiday exclusions. Calendar schedules use an IANA time zone; fixed intervals use elapsed time. The editor previews the next five occurrences.

For example, you could schedule a weekday workflow to inspect recent changes, run project checks, and summarize failures. This is an illustrative use case; configure prompts and permissions for your own project.

Loop after completion

Choose Execution → Repeat → Loop after completion, set a delay in seconds and either a run limit or Until stopped, then save and choose Start loop. The first iteration starts immediately. Each successful iteration starts a fresh session after the delay; iterations never overlap. Failure, cancellation or expiry ends the loop.

The loop keeps the version and permissions chosen at its start. Edits apply when you start a new loop. Run history labels each iteration; the workflow list and run details show progress, a countdown or a device waiting reason. Stop loop stops repetition and the current run, including during a delay. Archiving prevents future iterations while preserving the active run.

Paused or offline devices wait within the configured device waiting window from the next due time. Recovery starts at most one next iteration; overdue waits expire and revoked access ends the loop. By default, one inbox result is sent when the loop ends. You can also enable notifications for every successful iteration. Loop mode and timed schedules are separate choices.

Execution permissions

Ask for approval keeps native manual approvals for either agent. Approve for me is available for Codex using its native risk reviewer. Local and organization restrictions still apply. Agent questions and any forwarded approval requests may pause execution in every mode and count toward the total timeout.

Device availability and overlapping runs

Your computer must be awake and running a connected Caterm app or daemon to execute prompts. A run can wait for the device within its configured waiting window; it expires if the window is exceeded. There is no promise of execution while the computer is off.

Each workflow permits one active run. An overlapping scheduled occurrence is recorded as skipped. After server recovery, missed recurring cycles collapse to the most recent due occurrence within the waiting window; old prompts are not replayed as a backlog.

Context, history, and stopping

Each run starts a fresh conversation and native agent session. Steps within that run share context; later runs start separately. Editing a workflow changes future starts while active runs and loops keep their saved version and permissions. A failed or canceled step ends the run.

Pause a workflow to stop future timed triggers; manual runs remain available. Archiving stops future triggers and retains history. Neither action stops an active run: use its stop control. A daemon restart ends interrupted agent turns rather than automatically retrying their prompts.

Review remote execution boundaries →