Routines: scheduled agents that open pull requests

Some agent work should happen while nobody is watching: the CI digest that is waiting when you start the day, the dependency fixes that arrive as a pull request three times a week, the docs refresh every Monday. A routine is a saved prompt that Claude Code runs on its own, in the cloud, and ends with something a person reviews.

A Claude Code routine packages a prompt, one or more GitHub repositories, a cloud environment and a set of connectors, then runs it automatically. Each run is a full Claude Code session on Anthropic-managed infrastructure, so it keeps working when your laptop is closed. Routines are a research preview, available on the Pro, Max, Team and Enterprise plans.

A routine starts from one or more triggers:

• Schedule. Hourly, daily, on weekdays, weekly, a custom cron expression, or once at a set time. The minimum interval is one hour.
• API. An HTTP POST to the routine's own endpoint with a bearer token, for example from your alerting tool or deploy script.
• GitHub event. A pull request or release event on a repository, narrowed by filters such as base branch, labels or draft state.

One routine can combine all three. A review routine can run nightly, fire from a deploy script, and react to every new pull request.

Every run starts from a fresh clone. Claude reads the repository's CLAUDE.md and committed skills, does the work without stopping for permission prompts, pushes changes to a branch named claude/..., and leaves its output where the prompt says: a pull request, a comment, a Slack message through a connector. You open the run's session later to see exactly what it did.

Routines suit work that is unattended, repeatable and tied to a clear outcome.

• Digests. A daily CI health report, a Monday review-queue summary, a weekly list of stale feature flags.
• Maintenance with a check. Dependency and CVE sweeps, lint debt, docs that reference changed APIs. Each ends in a pull request with test output.
• Event-driven review. A team checklist applied to every new pull request, or a port of each merged change to a sister SDK.
• Alert triage. Your monitoring tool calls the API trigger with the alert text, and the routine correlates it with recent commits and opens a draft fix.

Choose something else when the job does not fit:

Do not schedule anything irreversible. A routine proposes. A person merges, deploys and sends.

Five things decide what a run can do.

1. The prompt. The trigger delivers your saved prompt as the session's task. Nobody is there to answer questions, so it must say what to do, where the output goes and what done looks like.
2. The repositories. Each one is cloned at the start of every run, starting from its default branch. A pull request trigger starts the first repository from the pull request's head commit.
3. The environment. It sets network access, environment variables and a setup script whose result is cached for later runs. The default environment allows package registries, GitHub, cloud SDKs and Anthropic's own sites, and blocks everything else.
4. The connectors. Every claude.ai connector you include is available with all of its tools, writes included, and no approval step. Remove the ones a routine does not need.
5. Your identity. A routine belongs to your claude.ai account. Its commits, pull requests and connector actions appear as you, and its runs count against your plan's usage.

Schedules are stored in UTC. The web form takes times in your local zone and converts them, and a custom cron expression goes in with /schedule update in the CLI. A run scheduled exactly on the hour can start several minutes late, so pick a few minutes past, such as 06:07. A one-off run fires once and then disables itself.

A green status in the run list means the session started and exited without an infrastructure error. It does not mean the task succeeded. Read the run.