Routines in Claude Code: How Anthropic turned Claude into a standalone developer that works without you

Anthropic has released one of the most long-awaited features for developers - Routines in Claude Code. This is not just another improvement, but a real breakthrough in agent coding. Now you configure a task once (prompt + repository + connectors), and Claude Code runs it itself: on a schedule, on an API request or in response to an event in GitHub. And all of this works on Anthropic's cloud infrastructure -- the laptop can be turned off and automation will live on.

The feature has been released in research preview, but is now available to all users of Pro, Max, Team and Enterprise plans with Claude Code enabled on the web. Let’s understand in detail: what it is, how it works, what scenarios the teams are already using and what nuances it is important to consider.

## What is “routine” and why is it more than a cron-job

Routines are the saved configuration of Claude Code

• Prompt (mainly - it should be as self-sufficient and explicit as possible).
• One or more repositories GitHub.
• Connectors (Slack, Linear, Datadog, Google Drive and other MCP connectors).

All this is packaged once and starts automatically. Claude Code clones a repository, runs a session in the cloud, performs a prompt, makes changes, opens a PR, writes to Slack – all without your input.

The main difference between the old /schedule in CLI is:

• Previously, you had to keep your own infrastructure, cron jobs, MCP servers.
• Now everything on Anthropic infrastructure is secure, scalable, independent of your laptop.

Routines can combine triggers: the same routine can run on a schedule, webhook, and API. Old scheduled tasks were automatically converted into routines upon release.

## Three Types of Triggers: How to Start Automation

**1. Scheduled routines ** Choose the frequency: every hour, every night, on weekdays, once a week. Time is specified in your time zone – the cloud will translate everything automatically.

Examples from the announcement and documentation:

• Every night at 2:00: pulls the top bug out of Linear, tries to fix and opens draft-PR.
• Weekly: scans merged PR, searches for “obsolete” documentation and opens PR with edits.
• Nightly: triage of new issues, label placement, owner appointment and summary at Slack.

2. By API (API routines) Each routine receives its own unique endpoint + bearer token. If you do POST, you get the session URL.

Perfect for integrations:

• Claude checks smoke tests, error logs and writes “go/no-go” to the channel.
• Claude correlates with recent commits, opens draft-PR with fix.
• Claude immediately opens the session and prepares changes.

*3. Webhook (only GitHub for now) * Subscribe to repository events (pull request.opened, push, issue, etc. + filters).

Cool cases:

• For every PR that touches the /auth-provider module: automatically summarizes the changes and posts to #auth-changes.
• On merged PR in Python SDK: Automatically port changes to the Go SDK and open the corresponding PR.
• Custom code review on the team checklist (security + performance + style) with inline comments before the person looks.

In the future, Anthropic promises to expand webhook to other sources of events.

## How to create a routine: step by step (web / CLI / Desktop)

** Through the web (claude.ai/code/routines):**

1. Give a name and write a prompt (choose a model).
2. Add repositories (with the option of unrestricted branch pushes).
3. Select environment (network access, env variables, setup script).
4. Configure the triggers (maybe several).
5. Disabling unnecessary connectors.
6. You do.

Through CLI: /schedule (for scheduled) + then edit the web. Through Desktop app: New remote task.

Important details:

• The routine is autonomous - no permission prompts or approval.
• Commits and PRs come from your GitHub account.
• Everything can be seen in individual sessions - you can monitor, review changes, bald.

## Limits and pricing (at the time of release)

• Pro: Up to 5 routines per day
• Max: up to 15
• Team/Enterprise: up to 25

Beyond the limit is extra usage. Each run of the routine spends tokens like a regular Claude Code session. In the research preview, behaviors and limits can change.

## Advantages and strategic implications

** For a solo developer:**

• Finally, you can sleep peacefully: backlog, docs, small bugs – everything is processed at night.
• There is no need to keep local scripts and servers.

** For the team:**

• Standardization of processes (code review checklist, library porting).
• Reduce the load on the on-call.
• Automatic synchronization between languages/repositories.

Broader implications:

• Claude Code is no longer just a “smart auto-addition” and becomes a real autonomous agent in your workflow.
• Shifting developer focus from routine to architecture, design and high-level solutions.
• Risk: If the prompt is bad, the routine will repeat errors at scale. Therefore, a quality self-contained prompt is 80% of success.

## Best Practices (from Documentation and Early Reviews)

• The prompt should be clear about the criteria for success.
• Limit the scope: only the necessary repositories, connectors and environment.
• Start with templates (Anthropic offers 8 ready-made ones).
• Test through "Run Now.".
• For complex tasks, break down into several small routines instead of one monstrous one.
• Monitor context and tokens – long sessions can degrade.

Anthropic clearly plans to expand triggers and capabilities. This is just the beginning of a fully autonomous dev cycle.

## Conclusion: Is it worth trying right now?

If you already use Claude Code, routines are must-try. Even a simple night cleaning backlog or automatic triage alert pays for the setup in the first days.

Click on claude.ai/code/routines or simply type /schedule into CLI.