Claude keeps working when your laptop is closed.
That is what a Claude routine does. It is a saved Claude Code setup that runs on Anthropic's cloud on a schedule, an API call or a GitHub event. This page says what it can do, where it stops, and how to make one you can trust.
What is a Claude routine?
A Claude routine is a saved Claude Code setup: a prompt, one or more repositories and a set of connectors. Anthropic runs it on its own cloud, so it keeps working when your laptop is closed.
It can start in three ways. On a schedule, such as every weekday morning. From an API call, when something outside Claude fires it. Or from a GitHub event, such as a pull request being opened.
A routine is good at work you would otherwise repeat by hand every week: reading new mail, watching a list of sources, writing a report, shortlisting applicants. You describe the job once. Claude does the reading and sorting, and you keep the decisions.
Routines and scheduled tasks are the same thing
Anthropic's docs now call them "scheduled tasks, formerly routines". If you searched for Claude scheduled tasks or Claude routines, you found the same feature under two names.
The old name still shows up in places, such as the claude.ai/code/routines address and the /routines shortcut for /schedule. In RoutineOS, "routine" means the thing you build: the folder, the Desk and the rules. The scheduled task is where it runs.
Anthropic's docs list scheduled tasks on paid Claude plans. Claude Code is the tested route for building and scheduling one. An install route through Claude Desktop or claude.ai exists, but it is not the one RoutineOS has tested for scheduling.
What a Claude routine cannot do
Know the edges before you rely on one. Each line says how we know it.
| Limit | What it means for you | How we know |
|---|---|---|
| Research preview | The platform says behavior, limits and the API surface may change. Something that works today can change upstream. | docs-stated |
| One hour minimum | A routine runs at most once an hour, in UTC. A faster schedule is rejected. | docs-stated |
| Fresh sandbox each run | Files from the last run are gone. A routine needs a database or a connector to remember anything. | pilot-verified |
| Cannot fire another routine | A run cannot start or edit other routines. To chain two, one writes and a later scheduled run reads. | pilot-verified |
| Locally installed plugins are not loaded | A cloud run does not load plugins you installed on your own machine. Skills committed to the repository do load. See Claude Code plugins. | docs-stated |
| Belongs to one account | A scheduled task belongs to one claude.ai account and is not shared with teammates. What it does through connectors appears as you. | docs-stated |
| Network works like an allowlist | Some hosts answer and others are blocked. Test every new host from a run before you depend on it. | pilot-verified |
A bare schedule is easy to start and hard to trust
A prompt on a schedule forgets what it found last week. It has no safe place to put results. It cannot ask your permission. And you cannot correct it without someone editing the prompt.
RoutineOS adds those parts, the same way every time. It is a Claude Code plugin, free (MIT), and needs a paid Claude plan.
Memory
A store keeps what each run found, so the next run knows.
A place to meet
The Desk is a private page where you see what was kept, what was dropped and why.
Permission
Approval gates stop anything that sends, posts or deletes until you say yes.
Steering
Typed settings you change on the Desk, without editing the prompt.
Improvement
It watches what you keep and drop, then suggests one change and shows why. It never applies it itself.
How to trust a Claude routine
Four habits. RoutineOS builds them in, and you can follow them without it.
- Make the first run report-onlyIt reads and reports. It sends nothing. You look at the output before it does anything else.
- Put a gate in front of every send, post and deleteThe routine drafts. A person says yes. Nothing leaves without it.
- Treat what it reads as information, not ordersIf a web page or an email says "do this", the routine reads it as data and never follows it.
- Call it done only when you have checked itScheduled is not finished. It is finished when it is live and a person has looked at a real run.
- Never sends, posts or deletes without your yes.
- Never applies its own suggestions.
- Never asks you to paste a secret into the chat.
Cost, how often it runs, and how it improves
The short answers are on the FAQ: what it costs, how often a routine can run, whether it remembers, and how it gets better. How RoutineOS compares with other tools is on Compare. The full docs start at docs.routineos.org.
Make your first routine
Install the plugin in Claude Code 2.1.275 or later. Ask for what you want in plain words, then check the first run.