Claude Code Routines are autonomous, recurring sessions that run on 📝Anthropic's cloud infrastructure — no laptop required. A routine packages a prompt, one or more repositories, connectors, and triggers into a unit of work that executes unattended.

How They Work

Each routine execution clones the specified repositories, creates a fresh 📝Claude Code session, and runs the prompt with full agent capabilities: file editing, shell commands, tool use, and access to any skills committed to the repo. No permission prompts fire during execution. Commits and PRs appear under your GitHub identity; Slack messages appear as you.

Trigger Types

Routines support three trigger types, combinable on a single routine:

Schedule

Run on a recurring cadence — hourly, daily, weekdays, weekly, or custom cron expressions (minimum interval: 1 hour). Times are in your local timezone. Runs may start a few minutes after the scheduled time due to stagger.

API

Each routine gets a unique /fire endpoint authenticated with a per-routine bearer token. Send an optional text field with run-specific context (alert body, error logs, deploy metadata). Returns a session ID and URL immediately. Requires the beta header experimental-cc-routine-2026-04-01.

GitHub

Subscribe to repository events — PR opened/closed/assigned/labeled/synchronized, or release created/published/edited/deleted. Filter conditions narrow which events trigger: author, title, body, base branch, head branch, labels, draft status, merged status. Requires the Claude GitHub App installed on the repository.

Configuration

Three surfaces for creating and managing routines, all synced to the same cloud account:

Web — claude.ai/code/routines (full UI)
CLI — /schedule command (conversational creation, listing, updating, running)
Desktop app — New task → New remote task

Repository Settings

By default, routines can only push to branches prefixed with claude/
Enable "Allow unrestricted branch pushes" per repository to remove this restriction
Clones start from the default branch unless the prompt specifies otherwise

Environment

Each routine runs in a cloud environment with configurable:

Network access — restricted, limited, or full internet
Environment variables — API keys, tokens, secrets
Setup scripts — cached so dependencies install once, not on every run

Connectors

All connected 📝MCP connectors (Slack, Linear, Google Drive, etc.) are included by default. Remove any the routine doesn't need to limit Claude's access scope.

Usage and Limits

Routines count against daily subscription usage (same pool as interactive sessions). There's an additional daily cap on routine runs per account, visible at claude.ai/code/routines. Organizations with extra usage enabled get metered overage; otherwise runs are rejected until the window resets.

Routines vs. Other Automation

Routines — cloud-native, unattended, triggered by schedule/API/GitHub events
Hooks — local, reactive, fire on pre/post tool-use events within a session
Skills — reusable prompt workflows invoked with slash commands
/loop — recurring tasks within an open CLI session (requires laptop)
Desktop scheduled tasks — local recurring work with file/system access

Example Use Cases

Backlog maintenance — nightly schedule → read issues → auto-label, assign, post summary to Slack
Alert triage — API trigger on monitoring threshold → correlate stack trace with commits → open draft fix PR
Code review — GitHub trigger on PR opened → apply team checklist → leave inline comments
Deploy verification — API trigger post-deploy → run smoke tests → scan logs → post go/no-go to Slack
Docs drift — weekly schedule → scan merged PRs → flag outdated docs → open update PRs

Status

Research preview as of April 2026. Available on Pro, Max, Team, and Enterprise plans with Claude Code on the web enabled. Behavior, limits, and API surface may change.