Claude Routines vs Desktop Scheduled Tasks vs /loop: Which Claude Automation Mode Should You Use?

Anthropic now gives Claude Code users three distinct ways to automate work: Routines, Desktop Scheduled Tasks, and /loop. They sound similar, but they are not interchangeable. One runs in Anthropic’s cloud, one runs on your machine, and one lives only inside your current session. If you pick the wrong one, you either lose local access, lose persistence, or end up overengineering a job that only needed a quick polling loop.

This guide is the practical decision layer. It explains what each automation mode does, where it runs, what it can access, when to use it, what it costs, and where competitors like Cursor Cloud Agents and GitHub Actions fit instead. It also pulls in early operator reaction, because the launch story is not just convenience. It is also about tradeoffs: portability versus convenience, cloud durability versus local control, and agent automation versus plain old deterministic workflows.

If you are new to Claude’s automation ecosystem, it helps to understand how Anthropic has been building toward this moment. Their Managed Agents launch in early April 2026 laid the infrastructure groundwork, and Routines are the scheduling layer that sits on top of it.

The key difference is not just scheduling. It is execution environment. Routines are cloud sessions that can keep working when you close the browser. Desktop Scheduled Tasks are local recurring sessions with access to your actual machine. /loop is the lightweight option: fast, temporary, and tied to the session you already have open.

What are Claude Routines, Desktop Scheduled Tasks, and /loop?

Claude Routines are saved Claude Code automations that combine a prompt, one or more repositories, and connectors, then run on a schedule, via API call, or from GitHub events. They run on Anthropic-managed cloud infrastructure, which is the whole point: the job keeps going even if your laptop is closed. Anthropic positions them as a way to stop duct-taping cron jobs, local infra, and extra tooling around Claude Code.

Desktop Scheduled Tasks live inside Claude Desktop. They start new sessions automatically at the time and frequency you choose, and they have direct access to your local files and local tools. That makes them the right choice for workflows that need your actual working directory, local MCP configuration, or uncommitted state. The tradeoff is obvious: if the app is not open, or the computer is asleep, the job does not run.

/loop is the quickest automation primitive. It re-runs a prompt on repeat inside the current Claude Code session. It can poll a deployment, babysit a PR, or keep checking a long-running build. But it is session-scoped, capped at 50 tasks per session, expires after 7 days, and disappears when you close the terminal or restart Claude Code.

Claude Routines vs Desktop Scheduled Tasks vs /loop at a glance

Which Claude automation mode should you use for your workflow?

This is where most explainers stop too early. The practical decision is not “Which one is best?” It is “Where does the work need to live?” If the task depends on cloud durability and event triggers, Routines win. If it depends on your machine’s state, Desktop Scheduled Tasks win. If it depends on momentum inside an already-open session, /loop wins.

When should you use Claude Routines?

Use Routines when the workflow should run without your machine, when triggers should come from outside your local session, or when the job is naturally repo-centric. Anthropic’s own examples are telling: nightly issue triage, docs drift detection, deploy verification, alert triage, feedback resolution, library porting across repos, and team-specific PR review. In other words, Routines are best when Claude needs to behave more like a cloud worker than a local assistant.

The biggest Routine caveat is also the easiest to miss: it does not run against your local environment. It runs in an Anthropic-managed VM with a fresh repository clone. Anything committed to the repo is available. Anything configured only on your laptop is not. That means local data, local-only tools, or uncommitted changes break the mental model if you choose Routines too casually.

For teams thinking about building AI agent workflows at scale — not just scheduling — our deep dive on 21 ways to make money with AI agents covers how these automation primitives feed into real business models.

When should you use Claude Desktop Scheduled Tasks?

Use Desktop Scheduled Tasks when local context matters more than cloud durability. That includes jobs like reviewing today’s local working directory, running checks that depend on your installed tools, or working with local MCP servers and configuration. They can run every minute, they survive restarts, and they support configurable permission prompts. That combination makes them far better than Routines for machine-bound automation.

The catch is operational, not conceptual. The Claude Desktop app must be open and the computer must be awake. If the machine sleeps, the scheduled run is skipped. Claude Desktop does support a catch-up behavior, but only for the most recent missed run in the last seven days. If timing matters, you need guardrails in the prompt itself. That is a smart detail Anthropic documents and one many comparison posts will ignore.

When should you use /loop in Claude Code?

Use /loop when you want Claude to keep checking something while you continue working in the same session. Anthropic explicitly calls out deployment polling, PR babysitting, long-running build checks, and reminders. This is the right tool for “keep an eye on this for me,” not “build an always-on automation.”

The constraints are sharp. A session can hold up to 50 scheduled tasks. Recurring tasks expire after seven days. There is no persistence across restarts, no catch-up across missed fires, and if Claude is busy when the interval hits, the task runs once when Claude becomes idle rather than replaying every missed interval. That is why /loop is great for live operator work and bad for durable automation.

If you are trying to estimate how much token budget a /loop task will burn through during a session, the Claude Code token calculator can help you model that before you commit to a long run.

What do Claude Routines cost, and what are the limits?

Claude Routines are available on Pro, Max, Team, and Enterprise plans. Anthropic’s launch post says Routines also have daily run limits: Pro gets 5 runs per day, Max gets 15, and Team or Enterprise gets 25. Extra usage can be enabled above those limits. That matters because many people will treat Routines like general-purpose automation, but Anthropic is still clearly constraining them as a preview feature with subscription guardrails.

Desktop Scheduled Tasks and /loop do not have published per-day run caps in the same way, but their limits show up elsewhere. Desktop tasks depend on your machine and app availability. /loop depends on the session and expires after seven days. In other words, Claude’s three automation modes each throttle you differently: subscription cap, machine availability, or session lifetime.

If you hit your Claude message limit while running Routines, be aware that the daily Routine cap and message rate limit are separate constraints — hitting one does not reset the other.

How do Claude Routines compare with Cursor Cloud Agents and GitHub Actions?

For comparison shopping, Claude Routines sit between a hosted coding agent and a workflow engine. The closest conceptual match is Cursor Cloud Agents—formerly called Background Agents—which also run in isolated cloud environments, clone your repo, can work in parallel, and are accessible from web, desktop, Slack, GitHub, Linear, and API. But Cursor charges Cloud Agents at API pricing and requires on-demand usage to be enabled, so the cost model is much closer to metered agent execution than subscription-bundled scheduling.

GitHub Actions is a different category. It is still the better fit when you need deterministic CI/CD-style automation instead of an autonomous coding agent. Pricing is based on runner minutes and storage. For example, GitHub lists Linux hosted runner pricing starting at $0.002 to $0.006 per minute depending on runner type, with standard hosted runners free for public repositories. That makes GitHub Actions cheaper and more predictable for fixed scripts, but much less ergonomic for “read the repo, reason about the situation, and take the next best action.”

What are the best use cases for each Claude automation mode?

The practical pattern here is simple. If the job is event-driven and repo-driven, Routines fit. If it is machine-driven, Desktop Scheduled Tasks fit. If it is attention-driven, /loop fits. That framing is much easier to use than memorizing feature checklists.

For a deeper look at how Anthropic thinks about delegation — including how to get Opus-level reasoning out of Sonnet-class automation — the Advisor Strategy guide is a useful companion read.

What are people saying about Claude Routines so far?

The early operator reaction is split in a useful way. The positive read is obvious: Anthropic finally gives Claude Code a cloud-side automation layer instead of forcing users to build their own cron wrappers, local infrastructure, or trigger plumbing. That is why launch coverage from 9to5Mac and The Decoder frames the story as cloud automation, autopilot bug fixing, and always-on PR review.

The skeptical read is just as important. In the Hacker News discussion around the launch, recurring themes include distrust that Anthropic will keep feature behavior stable, concern over vendor lock-in, frustration with provider-specific workflow state, and a strong preference for automation that can be rebuilt outside the vendor’s stack if needed. That is not a fringe complaint. It is a real operator instinct: convenience is attractive, but only if portability remains plausible.

What about the weird session-reuse detail in GitHub Routines?

There is one operator-level detail worth calling out because it affects how you think about GitHub-triggered automations. Anthropic’s launch post says Claude “opens one session per PR and will continue to feed updates from that PR to the session,” which sounds like ongoing session reuse. But the Routines docs say each matching GitHub event starts its own session and that session reuse across events is not available. If you are designing around this today, trust the docs more than the marketing line.

Watch Claude Routines in action

FAQ: Claude Routines vs Desktop Scheduled Tasks vs /loop

Bottom line: which Claude automation mode should you use?

If you want the shortest answer possible, here it is. Use Routines for durable cloud automation. Use Desktop Scheduled Tasks for recurring local work. Use /loop for temporary monitoring in a live session. Most people will end up using all three. The trick is to stop treating them as substitutes and start treating them as three different layers in the same automation stack.

That is also the larger story behind this launch. Anthropic did not just add another scheduling toggle. It moved Claude Code further toward a hosted agent runtime. That is useful. It is also strategic. And if you build with that in mind, you will choose the right mode more often – and avoid the wrong assumptions about what Claude can access, how long it will persist, and where your automation logic should really live.