Claude scheduled tasks: loops, desktop, or routines?

Use /loop for quick polling during a Claude Code session, a local Desktop task for recurring work that needs your files and tools, or a cloud routine for work that should run reliably without your machine or start from an API call or GitHub event. The right choice depends on whether the work needs your machine and local files, how long the schedule must persist, and what should start it.

This guide covers Claude Code scheduling, not every Claude product surface. Whenever is a separate workflow platform, not a settings screen for Claude.

Three modes. Three different operating contracts.

Current official documentation, reviewed October 10, 2026. Routines are in research preview; check the linked docs for changes to behavior and limits.

CLI /loop

In the current conversation

Best for checking a build, watching a deployment or revisiting a PR while a Claude Code session is running.

• Where it runs Claude Code must be running and idle for a task to fire. Local files and tools come from the session.

• What persists CronCreate tasks restore on --resume or --continue, except expired recurring tasks and overdue one-shots. Self-paced loops do not restore. Tasks are session-scoped.

• Timing & limits Fixed schedules have one-minute granularity, each fixed-schedule recurring task runs late by a fixed delay of up to 30 minutes, and recurring tasks, self-paced loops included, expire after seven days. A busy session runs a due task once when idle, not once for every missed interval.

Read the /loop documentation

Desktop local tasks

On your machine

Best for a recurring review or briefing that needs your local working folder and installed tools.

• Where it runs The Desktop app must be open and the computer awake. An open manual conversation is not required.

• What persists Saved tasks persist across app restarts. When the app starts or the computer wakes, Desktop starts one catch-up run for the most recently missed time within the last seven days.

• Timing & limits The minimum interval is one minute, set by asking Claude in a Desktop session, since the schedule picker’s shortest preset is hourly. Starts are staggered by a few minutes, and a run is skipped if the previous one is still in progress. A task in Manual permission mode stalls on a tool it lacks permission for until you approve it.

Read the Desktop task documentation

Cloud routines

Independent of your laptop

Best for recurring or event-driven Claude Code work using selected repositories, an environment and connectors.

• Where it runs Runs on Anthropic-managed infrastructure by default, or an organization’s configured self-hosted environment. No running laptop session is needed.

• What persists The configuration persists in your account; a routine skips runs while its GitHub connection is missing, turns off after 72 hours without one, and is on hold while your subscription is paused. Runs use fresh repository clones, not your laptop’s uncommitted files.

• Timing & limits Recurring schedules have a one-hour minimum, and a run scheduled exactly on the hour can start several minutes late. API triggers are also supported, and GitHub triggers once the Claude GitHub App is installed on the repository. Subscription usage and hourly run limits, per account and per routine, still apply.

Read the routines documentation

Set it up in the mode you actually need.

Start with a read-only task, and observe a real run before widening its access or leaving it unattended.

Poll in the CLI

/loop 5m check whether the deployment finished and report its status; do not change anything

1. Open Claude Code in the relevant project and verify the tools it needs are available.

2. Use an explicit interval for a fixed schedule; a prompt without an interval lets Claude choose the delay.

3. Ask “what scheduled tasks do I have?” to confirm it, and “cancel the deploy check job” when finished. Tasks stop when the session exits and recurring ones expire after seven days; for durable scheduling, use a Desktop task or a routine.

Schedule a local Desktop task

Code → Routines → New routine → Local

1. Choose a name, instructions, working folder, model, permission mode and schedule. Use an isolated worktree if the task should not work on your current uncommitted changes.

2. Choose Run now, watch for permission prompts and choose “always allow” for the tools this task needs, so later runs approve them automatically; allow rules in ~/.claude/settings.json also apply. MCP tools marked requiresUserInteraction still prompt on every call.

3. Keep the app open and the computer awake. Write prompts that tolerate a late catch-up run. Pausing the task stops scheduled runs without deleting it.

Create a cloud routine

/schedule daily PR review at 9am

1. Routines are available on Pro, Max, Team and Enterprise plans, and a Team or Enterprise Owner can turn them off for the organization. Create one from the routines web page, Desktop’s Cloud option, or /schedule in a CLI signed in with your claude.ai subscription; /schedule is not available with a Console API key or a cloud-provider login.

2. Set a self-contained prompt, repositories, environment and trigger. Review the connectors: currently all connected connectors are included by default, so remove the ones this routine does not need.

3. Run now, open the session transcript and inspect what actually happened. Check network access, outputs, account limits and write permissions before relying on the schedule.

Open Claude’s routines page

Scheduled does not mean safe to leave unchecked.

Check the outcome, not only the status

Claude’s routine docs say a green status means the session started and exited without an infrastructure error, not that the task succeeded. Read the transcript and verify the result.

Scope the authority

A routine can run shell commands and use included connectors, including writes, generally without pausing for approval. Limit repositories, network access and connectors to the job; some artifact actions still require approval.

Plan for timing and usage limits

These are not exact-second schedulers. Routines draw down subscription usage, and starting runs is also limited per hour: scheduled runs, one-off runs included, share 100 per hour per account. Usage credits can permit metered overage on subscription usage; the hourly limits have no overage.

Treat repeat actions as a design problem

Before posting messages or creating records, decide how the task identifies work it already processed: a late catch-up run, a manual Run now or a second trigger can start it again.

When the work has settled into a process.

Claude routines already offer cloud scheduling, as well as API and GitHub triggers. Choose Whenever for a different reason: when you want the recurring business process expressed as a typed workflow with explicit branches and integration calls, rather than a saved prompt, repositories and connectors that Claude carries out in a new session on each run.

For example: read yesterday’s onboarding rows, validate required fields, group blocked accounts by owner, then post one Slack summary. Those steps can be explicit code. Add AI only if one step needs classification or drafting.

Describe the rule

Brief

Name inputs, exceptions and the destination.

Inspect the logic

Workflow

Keep branches and AI boundaries explicit.

Check real inputs

Test

Include missing data and repeat delivery.

Run on a trigger

Publish

Review outputs and own the exceptions.

A separate path for recurring business workflows, not a replacement Claude scheduling interface.

Describe the process and Whenever’s AI writes the workflow, or point your coding agent at Whenever’s MCP server and have it build, test and deploy the workflow for you. This is not an automatic import of a Claude routine. Connections, permissions and runtime-specific behavior need review.

A few useful answers

Can Claude scheduled tasks run when my laptop is off?

Cloud routines can: they run on Anthropic-managed infrastructure by default, or a configured organization self-hosted environment. Local Desktop tasks need the app open and computer awake. CLI scheduled tasks need a running Claude Code session.

Does /loop survive a restart?

Scheduled tasks that haven’t expired are restored when you resume the same conversation with --resume or --continue. Expired recurring tasks and overdue one-shot tasks are not; self-paced loops, background Bash and monitor tasks are not restored. Tasks are session-scoped: closing the terminal or letting the session exit stops them firing, but backgrounding the session carries /loop tasks over to a background session that keeps running without a terminal.

What happens if a Desktop task misses its schedule?

On app startup or wake, Desktop checks the last seven days and starts one catch-up run for the most recently missed time. It does not replay every missed interval. The task may run much later than intended, so time-sensitive instructions need a guard.

Why can’t I schedule a cloud routine every five minutes?

The documented minimum interval for a cloud routine’s schedule trigger is one hour, and more frequent cron expressions are rejected. An API trigger can start a routine more often, within hourly limits: 30 starts per routine, shared by Run now, API fires and re-running a one-off, and 100 API fires per account. /loop (session-scoped scheduling in Claude Code) and Desktop local tasks have a one-minute minimum interval, but each needs its local runtime to stay available. Choose the mode for its full operating constraints, not the interval alone.

When would I use Whenever instead of a Claude routine?

Use a Claude routine when you want Claude to carry out recurring or event-triggered work across your repositories and connectors, such as grooming a backlog, reviewing PRs or triaging an alert. Consider Whenever when the business process has known steps you want represented as a typed workflow, with bounded AI only where it helps. Both can run without your laptop; the distinction is the execution contract, not a blanket cloud advantage.

Sources & further reading

Sources reviewed September 24, 2026.

• Claude Code: /loop, persistence, timing and limitations

• Claude Code Desktop: local tasks, setup, permissions and missed runs

• Claude Code routines: cloud setup, triggers, usage and constraints

• Whenever: typed workflows and trigger builders

Choose your next step

• Agent, workflow or hybrid?

• Build a typed workflow

• Build it from your coding agent

• Deploy with your coding agent

• Read Whenever’s workflow contract