Safeguard
Developer Tools•

Where do Claude scheduled tasks run: Cowork, Code, Desktop or cloud?

A map of the four places Anthropic's Claude products run scheduled work (Cowork, routines, Desktop, loops), what each can reach and when to use it.

Safeguard Security Team
6 min read

Claude has four places for scheduled work: Cowork scheduled tasks, Claude Code cloud routines, Claude Code Desktop tasks and session loops. They differ in where they run, whether your computer must be on, which files they can see and how they handle permissions. Pick by the job's needs, not by habit.

The four places

Cowork scheduled tasks. The Cowork help article says schedules can be hourly, daily, weekly, on weekdays or manual, and you can describe timing in plain language. Tasks run remotely, so they keep their schedule when your computer is asleep or the Desktop app is closed. A task that needs local files or apps runs only locally. They use the same connectors, skills and plugins as regular Cowork tasks, and each run is its own session. They are available on all paid plans.

Claude Code cloud routines. A routine packages a prompt, repositories and connectors and runs on Anthropic managed cloud infrastructure, or on a self-hosted environment when routed there. Triggers are schedule, API and GitHub events. Routines are in research preview and are available on Pro, Max, Team and Enterprise plans. The minimum interval is one hour.

Claude Code Desktop tasks. These run on your machine, with access to local files and tools, but only while the app is open and the computer is awake. Each task has its own permission mode. Missed runs produce one catch-up run.

Session loops. The loop command runs a prompt on an interval inside an open terminal session. Tasks are session scoped, recurring tasks expire after seven days, and they fire only while Claude Code is running and idle.

Side by side

Cowork scheduledCloud routineDesktop taskSession loop
Runs onRemote, unless it needs local filesAnthropic cloud or self-hostedYour machineYour machine
Computer must be onNo, for remote tasksNoYes, app open and awakeYes, session open
Local filesOnly if the task runs locallyNo, fresh cloneYesYes
Permission handlingApproval mode settingNo prompts, runs autonomouslyPer task mode, stalls on unapproved toolsInherits the session
TriggersScheduleSchedule, API, GitHubScheduleInterval
Minimum intervalNot stated in the article I read1 hour1 minute checks, presets from hourly1 minute

Choosing

  1. Does it need local files or tools? Yes means Desktop task or loop. No means cloud or Cowork.
  2. Must it run when your computer is off? Yes means cloud routine or remote Cowork.
  3. Is it a document or office style job or a code job? Cowork suits document and connector work. Routines fit repository work, such as docs drift or review.
  4. Does it react to events? Only routines offer GitHub and API triggers.
  5. Is it a quick poll during a work session? A loop is the lightweight choice.

Worked example: three jobs, three places

A manager wants a Monday summary of her calendar and unread mail. It uses connectors and no local files, so a Cowork scheduled task fits and keeps running even if her laptop is closed.

A developer wants a nightly check that finds documentation referring to changed APIs and opens pull requests. It needs only the repository and connectors, so a cloud routine fits. He scopes it to the docs repository, removes unrelated connectors and relies on branch protection for review.

A data analyst wants a task that processes a spreadsheet saved on her laptop every weekday morning. It needs local files, so a Desktop task fits. She turns on the Keep computer awake setting and knows that if the laptop sleeps through the time, one catch-up run will happen later.

Mistakes people make when choosing

The first mistake is picking a Desktop task for something that matters to others. A teammate waiting for a Monday report does not care that your laptop was in a bag. If the output is a shared dependency, run it in a place that does not depend on one machine.

The second is picking a cloud routine for a job that needs a local file, then discovering the clone has nothing to read. Routines start from a fresh clone of the repositories you add, and the environment decides network access and variables.

The third is forgetting the plan. Availability varies by plan and platform, and the Cowork help article says web and mobile are in beta for Team and for Enterprise when an owner enables them. Check your own plan before designing around a feature.

The fourth is treating all four places as one thing in a policy. They are not. A single sentence such as "scheduled tasks must be reviewed" works, but the evidence for each place comes from a different screen, and the kill switch for each is different.

Governance angles

Each place has different admin controls. Team and Enterprise owners can disable routines and control Cowork access. Local tasks and loops need device level controls such as the environment variable that disables the cron scheduler. Identity differs too: routines act as the creator's linked accounts. Keep a register with a column for the place. See who a scheduled task acts as.

Try it yourself

Click through it at your own pace, or open it full screen.

Notice what the assistant can see and do once it is connected to a tool, and apply the same question to each of the four places above.

Where Safeguard fits

Safeguard's MCP Guardrail, policies and guardrails, and agent identity are GA capabilities relevant to any of these places. The Safeguard CLI can also write code, and the download page lists a Local agent and a Cloud agent as Available.

Quick check

Do Cowork scheduled tasks run when the app is closed?

Yes if the task runs remotely, per the help article. A task that needs local files or apps runs only locally.

Which place has no permission prompts?

Cloud routines. The routines page says there is no permission mode picker and the session runs without stopping for approval, apart from some artifact actions.

Which plans have routines?

Pro, Max, Team and Enterprise, per the routines page. Routines are in research preview, so limits may change.

Sources

Last reviewed: 19 October 2026.

Never miss an update

Weekly insights on software supply chain security, delivered to your inbox.

Self-healing security runs on Safeguard.

Your first fix PR is minutes away.

No sales call required, even your agent can complete the purchase over MCP.