Claude Code routines and scheduled tasks for recurring website jobs

A Claude Code routine is a saved task that runs by itself on a schedule, when an API call fires it, or when something happens on GitHub. Anthropic defines it as "a saved Claude Code configuration: a prompt, one or more repositories, and a set of connectors, packaged once and run automatically." Routines run on Anthropic's cloud infrastructure, so they keep working when your laptop is closed.

For a website, that means recurring jobs such as a weekly broken-link report, a monthly check for pages missing meta descriptions, or a smoke test after each deploy. Two facts should shape every routine you write. First, routines are a research preview as of October 2026: "Behavior, limits, and the API surface may change." Second, a routine runs with no permission prompts, and anything it does through GitHub or a connector appears as you. Design routines to report and open pull requests, not to change your live site.

Three ways to schedule Claude Code

Anthropic offers three scheduling options. They differ in where they run and whether anyone approves actions.

Source: Anthropic's routines documentation and scheduling comparison, as of October 2026. Anthropic's tip: use cloud routines "for work that should run reliably without your machine," Desktop tasks "when you need access to local files and tools," and /loop "for quick polling during a session," such as /loop 5m check the deploy.

How a routine works

Each run is a cloud session, the same kind described in our guide to Claude Code on the web. It clones your repositories fresh, runs in a cloud environment that controls network access and setup, does the work, and pushes changes to a branch prefixed with claude/ unless your prompt says otherwise. Each run appears as a session you can open to read what Claude did.

A routine can have one or more triggers:

• Schedule: hourly, daily, weekdays or weekly presets, or a one-off run at a set time. Custom intervals go through /schedule update in the terminal, but nothing more frequent than hourly. Anthropic notes a run set exactly on the hour "can start several minutes late," so pick a time like 9:07.
• API: an HTTP POST to a routine-specific endpoint with a bearer token. A deploy script or monitoring tool can fire it. The token "is shown once and cannot be retrieved later."
• GitHub: pull request or release events on a repository with the Claude GitHub App installed, with filters such as base branch or "is draft." Webhook events beyond the hourly caps are dropped.

Autonomy, identity and why scoping matters

This is the part to read twice. Anthropic's docs say routines "run autonomously as full Claude Code cloud sessions: there is no permission-mode picker, and the session runs shell commands, uses skills committed to the cloned repository, and calls any connectors you include, all without stopping for approval." The fired prompt also "can't act as approval or consent for actions during the run."

And on identity: routines "belong to your individual claude.ai account. They are not shared with teammates." Anything a routine does "appears as you: commits and pull requests carry your GitHub user, and Slack messages, Linear tickets, or other connector actions use your linked accounts."

So what a routine can reach is set by three things: the repositories you select, the environment's network access and variables, and the connectors you include. Anthropic's instruction is to "Scope each of those to what the routine actually needs." In practice:

• Remove connectors you don't need. When you create a routine, all your connected connectors are included by default, and Claude "can use every tool from an included connector, including writes, without asking."
• Protect your branches on GitHub. Use branch protection rules or rulesets so routine branches can't land on your main branch without a reviewed pull request.
• Keep secrets out of environment variables. They are "visible to anyone who uses the environment." Anthropic recommends network secrets for API keys. Our guide to keeping secrets away from Claude Code has the wider rules.
• Treat fired text as data. Text sent with an API call arrives wrapped and labeled as untrusted, and Claude won't follow instructions in it unless your saved prompt says to. Keep it that way for anything public-facing.
• Decide whose routine it is. Commits and connector actions carry the owner's identity, so choose the owner before you build it.

Team and Enterprise Owners can turn routines off for all members with a Routines toggle in the admin settings. When it is off, existing routines stop and nobody can create new ones.

Recurring website jobs that suit a routine

Good routines are, in Anthropic's words, "unattended, repeatable, and tied to a clear outcome." These examples are illustrative. Each ends in a report or a pull request, not a change to the live site.

A smoke check against a staging or live URL also needs that domain allowed in the environment's network access. For the link job, our step-by-step guide to finding and fixing broken links with Claude Code explains what to check. A routine just repeats that work on a timer.

Writing a routine prompt that works unattended

Nobody is there to answer questions during a run, so Anthropic says the prompt "must be self-contained and explicit about what to do and what success looks like." A short illustrative example:

Check every internal link in the site's source files on the default branch. For links that point to a page that no longer exists, fix only cases where the correct target is certain from the redirects file. Open one pull request titled "Weekly link fixes" with a table of every change. List uncertain cases in the pull request description without editing them. Do not modify templates, forms, tracking code or the redirects file.

Notice what it does: it defines scope, names what not to touch, says what output to produce, and leaves judgment calls to a person. Copy-ready prompts are in our Claude Code templates.

1. Create the routine with one repository, a minimal connector set and the Default environment.
2. Click Run now and open the run session. Anthropic warns that "A green status in the run list means the session started and exited without an infrastructure error. It does not mean the task in your prompt succeeded." Read the transcript.
3. Review the pull request the way you would review a person's work, including the rendered pages.
4. Only then turn on the schedule. Revisit the routine's runs each week for the first month.

Usage and limits

• Subscription usage: routines "draw down subscription usage the same way interactive sessions do." Our Claude Code pricing guide explains plan limits.
• Hourly caps, as of October 2026: 100 scheduled runs per hour per account; Run now, API fires and one-off re-runs share 30 per hour per routine; Run now and API fires each have 100 per hour per account. "None of these hourly limits has overage."
• Past your plan limit: organizations with usage credits turned on can keep running on metered overage; otherwise extra runs are rejected until the usage window resets.
• GitHub connection: if it expires, runs are skipped for up to 72 hours, and after that the routine turns off until you reconnect.
• Sign-in: /schedule requires a claude.ai subscription login. It is unavailable with a Console API key or a cloud provider such as Amazon Bedrock. API triggers are added on the web.
• Local MCP servers: servers added with claude mcp add on your machine don't appear as connectors. Add them at claude.ai/customize/connectors.

Routines, GitHub Actions or a script?

Routines are quick to set up and run without your computer, but they belong to one person, run without approval and may change while in preview. GitHub Actions runs Claude Code inside your repository's own pipeline, with configuration the whole team can see and review in git. Our guide to Claude Code in GitHub Actions for website QA covers that route. For batch jobs a developer runs by hand or from a server, scripted runs are covered in our guide to headless mode and scripting.

A reasonable rule: use a routine for personal or small-team reports, and move a job into GitHub Actions once the team depends on it.

Conclusion

Routines turn a good Claude Code prompt into a recurring job that runs while you sleep, which suits link checks, content audits and post-deploy smoke tests. Because they are a research preview, run without permission prompts and act under the owner's GitHub and connector identities, build them narrowly: one repository, only the connectors needed, branch protection on, and output limited to reports and pull requests a person reviews. If you want help designing recurring website jobs with safe scopes and a review step, our Claude Code consulting covers that setup. More guides are in our Claude Code Guide.

Frequently asked questions