Operate Soft · unattended cloud run research preview ⓘ legend

Routines

A saved Claude Code configuration — a prompt + one or more repositories + connectors/MCP, packaged once and run automatically on Anthropic-managed cloud infrastructure, with your laptop closed. Each routine fires on one or more triggers: a schedule, a GitHub event, or an API call. Part of Claude Code on the web.

Analogy — a cron job with a brain, run in someone else’s data center. You write the standing order and hand over the keys (which repos, which connectors); the cloud wakes it on a timer, a repo event, or a webhook, and it does the whole job unattended — with no one at the keyboard to approve anything.

“Routines” vs “a routine” — the name collision If you’ve called a saved procedure “a routine,” you meant a skill — that’s still true. Anthropic now also ships a feature literally named Routines, but it’s a different thing: unattended cloud automation triggered on a schedule / GitHub event / API call, not a procedure loaded into a session. A skill = what to do; a routine = when & where it runs unattended.

Mechanism — a saved config + triggers

A routine packages a prompt, one or more repositories, an environment, and connectors into one saved object on your claude.ai account. A trigger then runs it as a full autonomous cloud session:

save once
the config
prompt + repos + environment + connectors, stored in the cloud

a trigger fires
schedule / GitHub / API
a timer, a repo event, or an authenticated POST

unattended run
a cloud session
no permission prompts; pushes to claude/ branches; review after

# CLI: scheduled triggers only (API/GitHub triggers are added on the web)
/schedule daily PR review at 9am     # create a recurring routine, conversationally
/schedule in 2 weeks, remove the feature flag   # a one-off future run
/schedule list                       # see all routines
/schedule update                     # edit cadence / set a custom cron
/schedule run                        # fire a one-off run now

# API trigger: each routine exposes its own /fire endpoint (illustrative)
curl -X POST https://api.anthropic.com/.../routines/<routine-id>/fire \
  -H "Authorization: Bearer $ROUTINE_TOKEN" \
  -H "anthropic-beta: experimental-cc-routine-2026-04-01" \   # experimental beta header
  -H "content-type: application/json" \
  -d '{"text":"optional run-specific context, e.g. an alert body"}'

Triggers, environment & branch reach

Triggers (combinable in one routine): Scheduled — recurring (hourly/daily/weekdays/weekly, min interval 1 h) or a one-off future run; custom cron via /schedule update. API — per-routine POST /fire with a bearer token + optional text. GitHub — pull_request.* / release.* with filters (author, base/head branch, labels, draft/merged). Each GitHub event starts its own session.

Environment: default Trusted network (package registries + common dev domains only); configurable network access, environment variables/secrets, and a cached setup script. Branches: by default a run can only push to claude/-prefixed branches unless Allow unrestricted branch pushes is enabled per repo.

Identity & limits: belongs to your individual claude.ai account; actions appear as you (your GitHub identity, your connectors). Draws down subscription usage like an interactive session, plus a per-account daily run cap (one-off runs exempt). Admins can disable routines org-wide.

Created/managed at claude.ai/code/routines, the Desktop app, or the CLI /schedule (CLI = scheduled triggers only; API/GitHub triggers are added on the web).

When to reach for it

✓

A maintenance job should run on a schedule with your laptop closed (dependency audit, changelog, backlog triage).

✓

Claude should react to a GitHub event (a new PR, a release) without you watching.

✓

An external system (monitoring, CD pipeline, internal tool) should fire a Claude run via an authenticated webhook.

When not to

A task that repeats inside an open session → /loop (in-session, local).

A schedule that must run on your own machine with local files → Desktop scheduled tasks.

A pure pass/fail CI gate that runs your scripts → GitHub Actions (not a Claude session).

Best practices

★

Scope tight: pick the minimum repos + connectors — an unattended run has no approval step to catch a wrong move.

★

Keep the default claude/-branch restriction on; only enable unrestricted pushes where you truly need it.

★

Verify behaviour with a one-off (/schedule run) before trusting the cadence — one-off runs are exempt from the daily cap.

★

Keep network on Trusted and pass secrets via the environment, not the prompt. Write a self-contained prompt — nobody’s there to clarify.

No prompts means no safety net A run is a fully autonomous cloud session: no permission prompts, no approval mode. Its blast radius is whatever you pre-selected — repos, branch-push setting, environment, connectors — and it acts as you. A green run status only means the session started and exited cleanly, not that the task succeeded — open the run to confirm. Scope every routine to the minimum it needs.

Sources

Routines · Claude Code on the web · /loop & in-session scheduling