Claude Code Routines Explained: Scheduled Cloud Agents for Business

Claude Code Routines are scheduled cloud agents that run a saved prompt on a schedule in Anthropic's cloud. Think of each one as a small cron job that runs a full AI agent, not a fixed script. They keep working when your laptop is closed.

This guide explains what routines are, how they work, and the recurring business jobs they handle well. We run many of these in production every day, so we fold in real operating detail. The goal is a clear decision, not hype.

What are Claude Code Routines?

Claude Code Routines are saved AI agent tasks that Anthropic runs automatically on a schedule, on demand, or on an event. Each routine bundles a prompt, one or more code repositories, an environment, and any connectors it needs.

The routine runs on Anthropic-managed cloud infrastructure. Your machine does not need to be on. That is the core difference from running Claude Code by hand in a terminal.

Anthropic launched routines in 2026. They are still labeled a research preview, so limits and behavior can change.

How do Claude Code Routines work?

A routine works by firing a trigger, which starts a fresh cloud session that runs your saved prompt end to end. The agent can read a repository, run shell commands, use skills, and call connectors like Slack or Linear.

There are three trigger types. A schedule trigger runs on a cadence like hourly, nightly, or weekly. An API trigger gives the routine an HTTP endpoint you POST to. A GitHub trigger fires on pull requests or releases.

You can combine triggers on one routine. The same prompt can run every morning, respond to a webhook, and react to a new pull request. You set them up in the Claude Code interface.

• **Schedule trigger:** recurring runs (hourly, daily, weekdays, weekly) or a one-off run at a set time.
• **API trigger:** a per-routine URL and bearer token you call from alerting or deploy tools.
• **GitHub trigger:** runs on pull request or release events, with filters by author, label, or branch.

How are routines different from a normal cron job?

A routine differs from a cron job because it runs a reasoning agent, not a fixed set of steps. A cron job does exactly what its script says. A routine reads context and decides what to do inside the boundaries you set.

For example, a nightly cron script might email a raw error count. A routine can read the logs, judge whether an error is a real bug or noise, correlate it with recent commits, and open a draft pull request with a fix.

The tradeoff is control. Cron output is deterministic and cheap. A routine is flexible but non-deterministic, so you review its work rather than trust it blindly. Treat the green run status as "it finished," not "it succeeded."

• **Cron job:** fixed steps, deterministic output, no judgment, near-zero cost.
• **Claude Code Routine:** reasons over context, adapts, can write code or summaries, needs review.
• **Both:** run unattended on a schedule without your machine on.

What business jobs work well as routines?

Routines fit recurring, repeatable work with a clear outcome that someone reviews. The best jobs run overnight or early morning and hand a person a finished draft instead of a blank page.

In our own daily operations, the reliable winners are read-and-summarize jobs and monitoring. A morning briefing that pulls yesterday's data and posts a summary rarely breaks. Jobs that write to production need tighter guardrails and human review.

• **Daily reports:** pull yesterday's metrics, sales, or search data and post a summary to Slack or email.
• **Monitoring:** watch a site, feed, or competitor and flag material changes.
• **Weekly audits:** scan for docs drift, broken links, or SEO issues and open review PRs.
• **Morning briefings:** a 7 a.m. digest of overnight activity so the team starts informed.
• **Backlog grooming:** label new issues, assign owners, and summarize the queue.
• **Alert triage:** an API trigger opens a draft fix when a monitoring alert fires.

What are the reliability gotchas?

The biggest gotcha is that a successful run status does not mean the task succeeded. A green check means the session started and exited without an infrastructure error. Blocked network calls and task failures still show green.

Network access is the second common trap. The default environment only reaches a trusted allowlist of domains. If your routine needs your own API or an outside site, add that domain to the environment first, or the run fails with a 403.

Two more we hit in production: runs start a few minutes late due to a fixed stagger, so schedule around soft deadlines, not hard ones. And routines act as you, so their commits, messages, and connector actions carry your identity.

• **Verify inside the run:** read the transcript, do not trust the status dot.
• **Whitelist domains:** the default environment blocks non-allowlisted hosts.
• **Write a self-contained prompt:** no human is there to answer follow-up questions.
• **Scope tightly:** include only the repositories, connectors, and branch access the job needs.

How much do routines cost, and what are the limits?

Routines are included with Claude Code on Pro, Max, Team, and Enterprise plans, with Claude Code on the web enabled. There is no separate routine fee. Each run draws down your subscription usage like a normal session.

On top of usage, each plan has a daily cap on how many routine runs can start. Paid Claude plans include a daily number of scheduled runs that scales with your plan tier. Routines is a newer, still-evolving feature, so exact limits vary by plan and can change. One-off runs typically do not count against that daily cap.

There is a minimum interval between scheduled runs, so very frequent schedules are not allowed. If you hit the cap or your usage limit, accounts with usage credits can keep running on metered overage; without credits, extra runs wait for the window to reset.

How do routines compare to workflow tools like n8n or Zapier?

Routines and workflow tools solve different halves of automation. n8n and Zapier connect apps with visual, rule-based steps. Claude Code Routines run an AI agent that reasons, writes code, and produces judgment-based output.

Use a workflow tool when the steps are fixed and predictable, like "new form entry, add a CRM row." Use a routine when the job needs reading, deciding, or drafting, like "review overnight errors and propose a fix."

Many businesses use both. A Zapier flow can catch an event and POST to a routine's API trigger, letting the deterministic layer hand judgment work to the agent.

• **n8n / Zapier:** best for fixed, multi-app, rule-based flows and high-frequency triggers.
• **Claude Code Routines:** best for reasoning, code changes, and drafting on a schedule.
• **Together:** workflow tool triggers the event, routine does the thinking.

How does a business get started with routines?

Start with one low-risk, read-only job that a person will review, like a daily summary posted to Slack. Write a self-contained prompt that states the task and what success looks like. Pick a small model and a tight environment.

Run it manually with Run now first. Read the full transcript to confirm the output is right before you trust the schedule. Then turn on the schedule and check the first few runs.

Once one routine is stable, add more one at a time. Keep each routine scoped to a single job so failures are easy to spot and fix.