1. Turn Routines on
Routines ships as an opt-in feature. Open Settings → Experimental and flip on Routines. A new Routines tab appears in Workspace, next to Chat and Code.
2. Decide chat or Code
A chat routine runs through a persona from Customize → Agents — the normal chat tool-loop, good for research, summaries, and anything that only needs web search or a page fetch. A Code routine runs a coding agent against a real workspace folder — pick this whenever the task needs to read or edit files, or run a command.
3. Create the routine
From the Routines tab, click New routine. Pick a flavor, describe the task in plain language, choose a model, and set a schedule — interval, daily, or weekly. For a Code routine, you'll also point it at a workspace path and pick which coding agent runs it.
You don't have to use the form. Ask chat or a Code session directly — "check my repo's tests every 6 hours and summarize any failures" — and the agent sets up the routine for you, describing what it created.
4. Confirm it
Every new routine lands in pending confirmation and does nothing until you explicitly confirm it in the Routines panel — this is true even for one an agent created on your behalf. Open it, review the prompt and schedule, and click Confirm to arm it.
5. Let it run — and watch what happens
Once confirmed, the routine fires on its schedule automatically. Each fire shows up in its run
history with the full result or error. If a Code routine's task needs a decision mid-run — running
a shell command in ask mode, for instance — the run pauses and shows a live approval
card right there in the history, exactly like an interactive session. Nothing runs silently.
Managing routines afterward
Pause a routine to take it offline without losing its setup, resume it later, or use Run now to trigger an already-confirmed routine immediately instead of waiting for its next scheduled fire. Edit the prompt, schedule, model, or agent at any time.
Ideas to start with
- Daily digest: a chat routine that searches the web each morning and summarizes the day's AI/local-LLM news.
- Nightly test run: a Code routine that runs your test suite on a schedule and reports failures.
- Dependency check: a weekly Code routine that checks for outdated packages and opens a summary.
Routines is opt-in — nothing is scheduled, and nothing runs, until you turn it on and confirm a routine yourself. See the Routines docs for the full reference.
Try it
That one command starts the daemon and opens the UI at http://127.0.0.1:6996. From
there: Settings → Experimental → Routines, then Workspace → Routines → New routine.