In this guide
To run Claude Code on a recurring business task, write the task down as instructions it reads every session, give it only the tools and permissions that task needs, schedule it as a cloud or desktop scheduled task, and check every run against a short list of acceptance tests. Start read-only, and let it change systems only after the checks pass.
This guide is for teams that have tried Claude Code interactively and want one task to run the same way every week, whoever starts it. Feature details come from Anthropic’s Claude Code documentation; confirm availability for your plan.
Which business tasks suit Claude Code?
Claude Code works on files, commands and connected tools, so it suits tasks that already live in those places: checking a website after each release, turning exported data into a weekly report, cleaning a spreadsheet before an import, or keeping documentation in step with a product. Good candidates repeat on a schedule, have an output someone can inspect, and have a clear owner who acts on the result.
Poor candidates need judgment nobody has written down, or write into systems where a wrong change is expensive to undo. Those can still work, but only after the decision rules exist and a person reviews each output.
What does Claude Code need to run a task reliably?
- Instructions it reads every time. Claude Code reads a
CLAUDE.mdfile at the start of every session. Put the task’s goal, inputs, output format, rules and the checks a run must pass there, so the result does not depend on how someone phrases the request. - A packaged workflow. Skills package a repeatable workflow under a name the team can run, so a report or an audit runs the same way each time.
- Only the tools the task needs. MCP servers connect Claude Code to data sources and tools such as Google Drive, Jira or Slack. Connect the ones the task reads from, and add write access only where the task must change something.
- Checks that run automatically. Hooks run shell commands before or after Claude Code actions, for example a validation script after every file edit.
How do you schedule Claude Code?
There are four common ways to run Claude Code on a schedule. The main difference is where each one runs, which decides whether your computer has to be on.
| Option | Where it runs | Needs your computer awake? | Local files | Good for |
|---|---|---|---|---|
| Cloud scheduled tasks (formerly Routines) | Anthropic’s cloud; each run clones the repositories you select | No | Only the cloned repositories and connected tools | Tasks on hosted repositories and connected tools, at most once an hour |
| Desktop scheduled tasks | On your own machine, in a local folder | Yes, and the desktop app must be running | Yes | Tasks that need local files, local tools or a signed-in browser session |
/loop | Inside one open Claude Code session | Yes, and the session must stay open | Yes | Short polling while you work, such as watching a deploy |
| GitHub Actions (scheduled workflow) | GitHub-hosted runners | No | Only the repository checkout | Repository tasks on a cron schedule |
Each option has limits worth knowing before you rely on it. A desktop task that the computer sleeps through is skipped; when the computer wakes, the app starts one catch-up run for the most recent missed time. /loop tasks stop when the session closes, expire after seven days and do not catch up. Cloud scheduled tasks are a research preview on the Pro, Max, Ultra, Team and Enterprise plans. A GitHub Actions schedule needs the Claude GitHub App and an API key or OAuth token stored as a repository secret. Details from Anthropic’s documentation on desktop scheduled tasks, cloud scheduled tasks, /loop and GitHub Actions, checked on 2 November 2026.
What we saw in our own operations. We run our website quality checks and content routines as desktop scheduled tasks. On 20 and 22 October 2026 several runs did not start at their scheduled time. We added a watchdog task that runs 45 minutes after each slot and starts any missed run while its window is still open, and a daily health check that flags any routine that has not run on schedule. If a missed run would cost you a deadline, use a cloud scheduled task or add a check like these.
How do you keep a scheduled run safe?
Start with a read-only version that produces a report and changes nothing. Compare a few weeks of reports with what a person would have concluded, then allow writes one action at a time. Keep credentials out of instructions and outputs, and give the task its own access with the narrowest permissions the tools allow.
Decide in advance what the run does when something is unclear: stop and report, skip the item, or ask a named person. A scheduled task that guesses is harder to trust than one that leaves a clear note saying what it could not decide.
How do you know each run worked?
Write five to ten acceptance checks before the first scheduled run. For a weekly report: the date range is correct, every number traces to its source export, no section is empty without a stated reason, and the owner received it. Have every run write a short log with the inputs it used, the checks it passed and anything it skipped.
Review the log weekly for the first month. If a run passes its checks and still produces something nobody uses, change what the task produces.
What should a handover include?
The instructions file, the skill or prompt, the list of connected tools and their permissions, the acceptance checks, an example run log and a named owner. With those, someone else can restart, pause or change the task without the person who built it.
We run our own website checks and reporting this way: written instructions, scheduled runs, a log per run and a person who reviews anything the run could not decide.
Frequently asked questions
Does a scheduled task run when my computer is asleep?
A desktop scheduled task does not. It runs only while the desktop app is open and the computer is awake, and a run the computer sleeps through is skipped. When the computer wakes, the app starts one catch-up run for the most recent missed time. Cloud scheduled tasks and GitHub Actions schedules run without your computer.
Do you need a developer to run Claude Code on a schedule?
Someone needs to write the instructions, connect the tools and define the checks. After that, a non-developer owner can review the logs and decide what changes; the instructions file is plain text.
Is Claude Code only for writing software?
It is built for coding work, and it also handles tasks made of files, commands and connected tools, such as reports, data checks and documentation updates.
To set up one recurring task with shared instructions, tool access, review checks and a handover, see Claude Code implementation.
Reviewed by Andrii Gorbenko, founder of AGI Technologies