Claude Code Routines: The Trigger Layer That Makes Claude Always-On

Claude has been useful for a while. But it has always needed you to start the conversation.

You could wire it up to do remarkable things. Multi-step agents, complex reasoning pipelines, decisions that would have taken hours manually. What you could not do was make Claude run by itself on a reliable cadence, without writing custom cron jobs, webhook receivers, and glue code to hold it together.

That gap just closed. On April 14, 2026, Anthropic released Claude Code Routines in research preview.

What Routines Is

Routines is a cloud-native automation layer built directly into Claude Code. Not a third-party integration. Not a wrapper someone built on top of the API. Anthropic's own answer to the question: "How do I make Claude run without me in the loop every time?"

The mechanism is straightforward. You define a Routine: a task for Claude to execute, with a trigger attached. When the trigger fires, Claude runs the task. No manual session. No custom server. No bespoke scheduling infrastructure.

Three trigger types ship with the research preview: Schedule, API, and GitHub.

The Three Triggers

Schedule: Cron-Style Recurring Runs

Schedule triggers work on a time-based cadence. You define when you want the Routine to fire and Claude runs the task at that interval. Think cron, but instead of calling a shell script, you are calling Claude with a defined task.

This is the trigger type most teams will reach for first. Morning briefings, daily reports, end-of-week digests, overnight processing jobs. Any task that runs on a clock is a candidate.

One concrete use case: a daily competitive intel routine. Every morning at 6am, Claude pulls from defined sources, surfaces new developments, and writes a structured summary into your Slack channel or shared doc. You show up to that summary. You never set it up again.

The value is not just the automation. It is the reliability. Custom cron jobs calling the Claude API break when your server restarts, when your API key rotates, or when the script drifts. Schedule triggers live inside Claude Code infrastructure. That is Anthropic's problem, not yours.

API: HTTP-Triggered from Any System

API triggers turn Claude into a callable service. You get a Routines endpoint, POST to it from any system that can send an HTTP request, and Claude runs the defined task.

This is the trigger type that connects Routines to the rest of your stack. A new lead hits your CRM: POST to the Routine, Claude analyses the lead profile and drafts a personalised outreach. A customer submits a support ticket: POST to the Routine, Claude categorises, prioritises, and drafts a response. A new document uploads to your system: POST to the Routine, Claude extracts key data and populates a structured record.

Any system that can send a webhook can trigger a Claude Routine. That is most of them.

The distinction from using the Claude API directly matters here. Direct API calls are stateless and require you to manage the request-response cycle in your code. Routines add a defined task context, managed execution, and cloud-native reliability. You are not calling Claude as a function. You are triggering Claude as a service.

GitHub: Event-Driven from Repo Activity

GitHub triggers fire Claude based on what happens in your repositories. Pull requests opened, commits pushed to main, issues created or labelled, reviews requested. Any GitHub event you can define as a trigger becomes a potential Routine entrypoint.

This is the trigger type developers will find most immediately useful. The use cases are obvious once you see them.

PR opened: Claude reviews the diff, checks against your contribution guidelines, flags potential issues, and leaves a structured comment. You still approve. Claude does the first pass.

New issue created: Claude reads the issue, cross-references existing issues for duplicates, adds appropriate labels, and posts a triage summary. Your team works from triage output, not raw inboxes.

Push to main: Claude generates a plain-English changelog entry from the commit messages and files changed. Documentation stays current without anyone manually writing it.

One point worth being clear on: GitHub triggers are not GitHub Actions. GitHub Actions runs shell commands and scripts in a CI/CD pipeline. GitHub triggers in Routines fire Claude tasks. The two can coexist and often should. One handles your build, test, and deploy pipeline. The other handles reasoning tasks on top of your code activity.

The Power-Stack: Routines Plus Managed Agents

Routines get a lot more powerful when you pair them with the Managed Agents API.

Here is the composition. Routines handle the trigger layer. You define when or why Claude should start work. Managed Agents handle the execution layer. When the trigger fires, instead of a single Claude session doing all the work, Managed Agents spin up purpose-built agents, track their state, coordinate their output, and report results back.

Routines fire the gun. Managed Agents do the running. Anthropic's own engineering post on Managed Agents explains the decoupling between the harness and the execution environment, which is what makes this composition work.

The practical result: you can build Claude-centric automation workflows that are not just always-on, but genuinely scalable. A GitHub trigger fires when a large codebase change lands. Managed Agents fan out across the affected modules, each running parallel analysis. A consolidated report comes back. One trigger, many agents, structured output.

Before these two capabilities existed together, building that workflow meant maintaining your own orchestration infrastructure: task queues, agent state management, retry logic, result aggregation. Now Anthropic maintains it. You define the work.

What This Displaces

Be specific about this, because the answer matters for where you invest engineering time.

Custom cron jobs calling the Claude API. If you have a cron-based script that calls Claude on a schedule and does some processing, Routines replaces that. You get better reliability, zero infrastructure maintenance, and native integration with Claude Code tooling.

Simple self-hosted n8n workflows where Claude is the core logic. If you built an n8n workflow primarily to give Claude a scheduled trigger or a webhook entrypoint, and Claude is doing the actual reasoning work, Routines likely displaces that workflow. Why maintain a self-hosted n8n instance to wrap Claude when Claude now has its own native trigger layer?

Glue code that existed only to give Claude a trigger. Any custom code you wrote specifically to initiate Claude sessions on a schedule or in response to an event is a candidate for replacement. That code's job was infrastructure. Routines is that infrastructure.

What This Does Not Displace

n8n for multi-vendor workflow orchestration. If your n8n workflow connects Salesforce, HubSpot, Stripe, and Slack, with Claude as one step among many, n8n stays. Routines is a trigger layer for Claude-centric work. It is not a general-purpose integration platform.

GitHub Actions for CI/CD. Your build pipelines, test runners, and deployment workflows stay in GitHub Actions. Routines fires Claude tasks on repo events. It does not run your tests or manage your deployments.

Zapier for consumer-grade business automation. Zapier's strength is breadth of integrations and low-code accessibility. Routines requires you to be comfortable in Claude Code. Different tool, different audience.

The honest framing: Routines is a specialist tool for Claude-centric automation. If Claude is the core logic, Routines is worth evaluating. If Claude is one component in a broader multi-vendor pipeline, your existing orchestration layer probably stays.

What to Build with Routines First

If you have access to the research preview, three starting points are worth your attention.

Daily inbox triage. Schedule trigger, fires each morning. Claude works through a defined set of inputs (emails, tickets, Slack threads, whatever your team surfaces), categorises them by urgency and type, and produces a structured triage output. Your team starts the day from the triage, not the raw inbox.

PR review routine. GitHub trigger on pull request opened. Claude reads the diff, checks against your documented standards, and posts a structured review comment within minutes of the PR landing. Human reviewers make the judgement calls. Claude does the first-pass pattern matching.

Weekly competitive intel digest. Schedule trigger, fires Sunday evening. Claude aggregates from defined sources, surfaces notable developments from the past week, and writes a structured summary. Monday morning your team is already briefed.

All three are low-risk starting points. They support human decision-making rather than replacing it. They produce structured output you can verify. And they save real hours every week.

Where This Sits in the Stack

In the Five Levels of Claude unlock, Routines sits at Level 4: autonomous, event-driven operation.

Level 3 is where most teams are today. Claude Code running in interactive sessions, doing meaningful work, but still human-initiated. Level 4 is where Claude starts working without you starting it. Routines is the mechanism that gets you there.

Level 5 is full orchestration: Routines firing triggers, Managed Agents executing work, outputs feeding back into other Routines, the whole system running as infrastructure rather than as a tool you pick up and put down.

That is where this is heading. The Claude stack guide maps the full progression if you want to see how these capabilities layer.

Where This Leaves You

Claude has been able to do the work for a while. Now it can do the work without waiting for you.

Routines is in research preview as of April 2026. The API surface may shift, access may be limited, and you should not build hard dependencies on it in production yet. But the direction is clear. Anthropic is building the trigger layer. The always-on gap is closing.

If you are running a consulting practice, an agency, or an ops-heavy business on Claude Code today, the question is not whether to use Routines. The question is which tasks are already good candidates and how fast you can move when full availability lands.

---

If you want a map of how these capabilities fit into your specific operation, the IP Assessment is where to start. It takes ten minutes and gives you a clear read on where the opportunity is.

Take the IP Assessment

If you are already running AI infrastructure and want to work through what the Routines plus Managed Agents stack looks like for your specific situation, the Program walks through exactly how we structure a 90-day engagement. This is one piece of running AI as an operating system. The full picture is in our AI orchestration guide.

Explore the Program