On 14 April 2026, Anthropic opened the research preview of Claude Code Routines. This is the feature that a lot of people building on Claude Code have been waiting for: saved automations that run on Anthropic's cloud infrastructure instead of your local machine, triggered by a schedule, an API call, or a GitHub event.
If you have been building scheduled AI tasks the manual way (cron jobs, Python scripts, server-hosted pipelines), routines do not replace all of that. But they do collapse a significant chunk of the setup work for a specific category of automation. Here is what they actually are, what they can do today, and where the limits sit.
What changed on 14 April
Before routines, Claude Code ran on your machine. You could use /loop to poll a task repeatedly in an open session, or /schedule to run something at intervals, but both required your laptop to stay open and connected. Close the lid, lose the session.
Routines move execution to Anthropic's cloud. You configure a prompt, point it at one or more repositories, attach any connectors you need (Slack, Linear, Google Drive, GitHub), and the routine runs independently. Your laptop can be off. The automation keeps going.
Alongside routines, Anthropic also shipped a redesigned Claude Code desktop application with an integrated terminal, a faster diff viewer, an in-app file editor, and multi-session support. But the routines are the headline feature for anyone building production workflows.
What is a routine
A routine is a saved Claude Code configuration. Three components:
- A prompt. The instructions Claude follows when the routine fires. This needs to be self-contained because the routine has no access to external context beyond what you give it.
- One or more repositories. Claude clones these from the default branch at runtime. It works on a fresh copy every time, creating
claude/-prefixed branches for any changes. - Connectors. MCP connectors for external services: Slack, Linear, Google Drive, GitHub. These give the routine the ability to post messages, create issues, read documents, or interact with pull requests.
You set it up at claude.ai/code/routines or from the CLI with /schedule. Once created, you hit "Run now" to test, then let the triggers handle the rest.
The three trigger types
Every routine needs at least one trigger. You can combine multiple triggers on the same routine.
Schedule trigger
Fixed intervals: hourly, daily, weekdays, weekly, or a custom cron expression. The minimum interval is one hour (if you need per-minute polling, /loop in an active session is still the tool for that). Timezone is local, not UTC.
API trigger
Each routine gets a unique HTTP endpoint with a dedicated bearer token. You fire it with a POST request:
POST https://api.anthropic.com/v1/claude_code/routines/[routine-id]/fire
Authorization: Bearer [token]
Content-Type: application/json
Body: {"text": "optional payload"}
The optional text field appends to the configured prompt, so you can pass context from the triggering system. This is the integration point for alerting tools, CI/CD pipelines, or internal dashboards that need to kick off a Claude task programmatically.
GitHub trigger
Hooks into the Claude GitHub App webhook. Supported events include pull requests, pushes, issues, releases, check runs, workflow runs, discussions, and more. You can filter PR triggers by author, title, body, base/head branch, labels, draft state, and fork origin. This requires an explicit Claude GitHub App installation on your repository, not just CLI authentication.
Connectors and repositories
Connectors are MCP integrations that give the routine access to external services. The supported set at launch includes Slack, Linear, Google Drive, and GitHub. All configured connectors are included by default when a routine runs. Anthropic recommends removing unused connectors to reduce the access surface.
Repositories are cloned fresh from the default branch every time the routine fires. Claude works on a claude/-prefixed branch and can open pull requests, but cannot push directly to main without an explicit override. This is a sensible safety default for unattended automation.
Six real use cases
Anthropic documented six scenarios in the launch. These map closely to the kinds of scheduled and event-driven automations I have been building for clients with custom Python pipelines:
- Backlog maintenance. Nightly labelling and assigning of issues, with a summary posted to Slack each morning.
- Alert triage. When a monitoring alert fires, the routine correlates stack traces with recent commits and opens a draft PR with a suggested fix.
- Code review. On
pull_request.opened, the routine applies your internal review checklist and posts inline comments. - Deploy verification. After a deployment, the routine runs smoke checks and posts a verdict to the release channel.
- Documentation drift. Weekly scan of merged PRs to detect changes that need documentation updates, then auto-updates the docs repository.
- Library porting. When an SDK changes in one language, the routine auto-ports the changes across other language variants.
The common thread: these are tasks that follow a repeatable pattern, run on a trigger, and produce output that a human reviews rather than output that requires a human decision at every step.
Limits and pricing
Routines are available on Claude Code Pro, Max, Team, and Enterprise plans. Usage counts toward your existing subscriber limits, with daily caps:
- Pro: 5 routines per day
- Max: 15 routines per day
- Team / Enterprise: 25 routines per day
Exceeding daily limits incurs additional charges for organisations with extra usage enabled. Others see runs rejected until the daily reset. The minimum scheduled interval is one hour. There is no automatic retry on failure, but you can configure the routine prompt to send a Slack alert or open a Linear issue when errors are detected.
This is still a research preview. The API requires a beta header (experimental-cc-routine-2026-04-01), and Anthropic notes that request and response shapes may change during the preview phase.
Routines vs desktop tasks vs /loop
Routines are not a replacement for everything. Here is where each tool fits:
- Routines: unattended cloud execution, minimum one-hour interval, triggered by schedule, API, or GitHub event. No local file access (fresh clone only). Best for overnight jobs, CI/CD integration, and event-driven workflows.
- Desktop tasks: run on your machine without an open session. Per-minute intervals possible. Full local file access. Best for tasks that need your local environment or files not in a repository.
- /loop: runs in an active Claude Code session on your machine. Per-minute polling possible. Full local access. Best for quick iterative tasks while you are working.
The key distinction: routines are for fire-and-forget automation that should keep running whether you are at your desk or not. Desktop tasks and /loop are for work that benefits from local context or needs faster intervals.
What this means for business automation
For the kind of work I do at JQ AI SYSTEMS (custom Claude API pipelines, scheduled reporting, multi-agent systems), routines sit in an interesting middle ground. They are not a replacement for full custom Python pipelines. A routine cannot orchestrate multiple agents with shared memory across runs, manage a database, or handle complex branching logic that spans days. Those still need custom infrastructure.
What routines do well is handle the simpler, single-purpose automations that previously required you to set up a server, write a scheduler, and manage deployment just to run a Claude task on a timer. Deploy verification, code review automation, documentation maintenance, alert triage: these are tasks that previously needed more infrastructure than the task itself warranted.
For clients running smaller-scale automations, routines could replace what would otherwise be a custom build. For clients running larger systems, routines become a useful component: the lightweight trigger layer that kicks off heavier workflows or handles the simpler tasks in the pipeline.
The broader signal is clear. Anthropic is building toward a world where Claude does not just answer questions when asked. It runs continuously, reacts to events, and operates autonomously within defined boundaries. Routines are the first managed version of that pattern. Expect them to get more capable as the preview progresses.
If you are already building scheduled AI tasks or thinking about automating repetitive development workflows, routines are worth testing now. If you need something more complex (multi-agent orchestration, persistent memory, custom integrations beyond the current connector set), that is where a custom AI automation build still makes sense. And if you are not sure which category your workflow falls into, the free 30-minute consultation is the place to start.