Intermediate

Routines

Routines are persistent, cloud-hosted Claude Code agents that run automatically on a schedule, in response to an API call, or when a GitHub event fires — without you keeping a terminal open.

Research preview (April 2026): Routines launched 14 April 2026. Behaviour, daily limits, and the API surface may change while the feature is in preview.

What Is a Routine?

A routine is a saved Claude Code configuration consisting of three things: a prompt (the instructions Claude runs), one or more GitHub repositories to work in, and a set of connectors (MCP services the routine is allowed to use). Once configured, each trigger creates a fresh, isolated cloud session on Anthropic-managed infrastructure — Claude executes the task autonomously, with no approval prompts and no need for your machine to be on.

This is distinct from the session-local /loop skill, which runs inside your current terminal session and stops when you close it.

Three Trigger Types

A routine can have any combination of these triggers attached to it:

Scheduled trigger

Runs on a recurring cadence — hourly, daily, weekdays, or weekly — or once at a specific future time. Minimum interval is one hour. Your local timezone is converted automatically. One-off runs do not count against the daily routine cap.

API trigger

An HTTP POST to a per-routine endpoint fires a run immediately. The request body accepts an optional text field for run-specific context (e.g., a monitoring alert body or failing log snippet). The response returns a session URL and ID so you can watch the run live. Useful for wiring routines into CI/CD pipelines or monitoring tools.

GitHub trigger

Fires when a repository event occurs. Supported events: pull request lifecycle (opened, closed, assigned, labeled, synchronized), and releases (created, published, edited, deleted). Filters are available for author, branch, labels, draft status, and more. Requires the Claude GitHub App installed on the target repository.

Routines vs /loop

Aspect	Routines	/loop
Execution location	Anthropic cloud VM	Your local terminal session
Persistence	Runs when your machine is off	Stops when session closes
Minimum interval	1 hour	No minimum (can run every minute)
Local access	No — cloud only	Yes — full local filesystem
Best for	Set-and-forget automation	Active-session short-lived polling

Projects and Environments

Each routine runs inside an environment — a cloud runtime configuration that controls what the session can access. You configure environments from the web UI or the /remote-env CLI command.

What carries into the session

• CLAUDE.md, .claude/settings.json hooks, .mcp.json servers
• Skills, agents, and custom commands committed to the repo
• Plugins declared in the repo's settings
• Pre-installed runtimes: Python, Node, Go, Rust, Java, Docker, PostgreSQL, Redis, and more

What does NOT carry over

• User-level ~/.claude/CLAUDE.md
• MCP servers added locally with claude mcp add
• Interactive auth (e.g. AWS SSO) — no secrets store yet
• Static API tokens/credentials

Network access has four levels: None, Trusted (default — package registries, GitHub, and 100+ common dev domains), Full (any domain), or Custom. You can also define environment variables and a setup script that runs once and is cached for ~7 days, so dependencies are already installed at session start.

Creating a Routine

Two ways to create routines:

Web UI at claude.ai/code → Routines → New Routine: name it, write the prompt, pick repositories, select an environment, choose trigger type(s), review connectors, click Create.
CLI via /schedule: /schedule daily PR review at 9am (recurring) or /schedule clean up feature flag in one week (one-off). CLI creates scheduled triggers only — add API or GitHub triggers on the web after.

Managing Routines

From the web UI

• Run now — trigger a run immediately from the detail page
• Edit — change name, prompt, repositories, environment, or triggers
• Pause / Resume — toggle from the Repeats section without deleting
• Delete — removes the routine; past run sessions remain in session list
• Filter sidebar — filter by Status (Active / Archived), Project, Environment, Last activity

From the CLI

• /schedule list — view all routines
• /schedule update — modify an existing routine
• /schedule run — trigger a run now

Plan Limits

Pro

5 routine runs / day

Max

15 routine runs / day

Team / Enterprise

25 routine runs / day

One-off scheduled runs do not count against the daily cap. Routine runs draw down subscription usage the same as interactive sessions. Team/Enterprise plans with extra usage enabled can continue on metered overage after the cap is hit.

Common Use Cases

Scheduled — nightly

Backlog triage: read new issues, apply labels, assign owners, post summary to Slack

GitHub — PR opened

Bespoke code review: apply checklist, leave inline comments, add summary

API — from monitoring tool

Alert triage: pull stack trace, correlate with recent commits, open draft PR with proposed fix

API — from CI/CD pipeline

Deploy verification: run smoke checks, scan error logs, post go/no-go to release channel

Scheduled — weekly

Docs drift: scan merged PRs, flag outdated docs, open update PRs

GitHub — PR merged

Cross-repo port: port a change merged in one SDK to a parallel SDK in another language

Security and Isolation

Each routine run executes in an isolated, Anthropic-managed VM. All outbound traffic passes through Anthropic's proxy for abuse prevention. GitHub authentication uses a scoped credential proxy — your tokens are never inside the sandbox directly. By default, Claude can only push to claude/-prefixed branches; enable "Allow unrestricted branch pushes" per repository if you need Claude to push to other branches.

Because cloud sessions run from Anthropic's infrastructure, IP allowlisting on external services is incompatible with routines — the outbound IP is not your network's IP.

Checklist: Do You Understand This?

A routine runs on Anthropic cloud infrastructure — your machine does not need to be on
Three trigger types: scheduled (min 1 hour), API (HTTP POST), GitHub (repo events)
Routines vs /loop: routines persist; /loop is session-local and has no minimum interval
Environments control network access, env vars, and setup scripts — cached for ~7 days
Daily caps: Pro 5, Max 15, Team/Enterprise 25 — one-off runs are exempt from the cap
Manage via the web UI (Routines sidebar with Status/Project/Environment filters) or CLI (/schedule commands)