A Claude routine is a Claude Code configuration you save once, then let run on its own: a prompt, one or more repositories and a set of connectors. It runs on Anthropic's servers, so your laptop can stay closed. It starts on a schedule, on an API call or on a GitHub event. Anthropic has since renamed the feature to scheduled tasks, and the official guide still describes it as a research preview, checked in November 2026.
The real question for a company comes right after. Does this replace n8n? Tandem deploys both, across around 100 automation engagements. The answer fits in one sentence. Claude scheduled tasks take over your scripts and your small repository chores. Your business automation platform keeps its ground.
What exactly is a Claude routine?
A Claude routine packages everything a session needs to work without you. The prompt describes the assignment and what success looks like. The repositories you select are cloned fresh at the start of every run. The connectors open access to your tools, Slack, Linear, a CRM, a documentation space. The environment sets the level of network access and the variables available.
Each run is a full Claude Code session. It runs shell commands, uses the skills committed to the cloned repository and calls the connectors you allowed. Permissions default to auto mode, where a classifier model allows or blocks the actions that would normally wait for your approval. A task stays attached to your individual account, it is not shared with teammates, and everything it does carries your GitHub identity.
The name changed, the feature stayed the same. Anthropic now says scheduled tasks, but the word routines survives in the web address, in the /routines command alias, in the API path and in the organisation settings. If you are hunting for the feature and your interface shows one or the other, it is the same thing.
It is available on the Pro, Max, Ultra, Team and Enterprise plans. You create one from the web, from the desktop app, or by typing the scheduling command inside a session. On Team and Enterprise, an owner can switch the feature off for the whole organisation, and existing tasks then stop running.
The three triggers: schedule, API and GitHub event
The schedule trigger covers the most common case, a recurring cadence or a single run at a future date. Presets go from hourly to weekly, and a custom cron expression handles the rest. Mind the floor, an expression that runs more often than once an hour is rejected.
The API trigger turns the task into an HTTP endpoint. You send a POST to its own address with a bearer token, and the session starts. The token is shown only once at creation, so copy it before closing the dialog. This is the trigger that wires a Claude task into your monitoring tool, your deployment pipeline or an internal form.
The GitHub trigger reacts to repository events, a pull request opened, a merge, a release. Each event opens an independent session, so two successive updates produce two separate sessions. Hourly caps apply per task and per account, and events beyond them are dropped until the window resets.
What a Claude scheduled task genuinely does well
Good use cases have one thing in common. The work is repeatable, it is judged on a clear outcome, and nobody wants to do it by hand on a Monday morning. Anthropic lists several examples in its documentation, and they overlap with what our clients ask for first.
- Backlog grooming: a nightly pass reads the issues opened since the previous run, applies labels, assigns owners based on the area of code involved, then posts a summary into a team channel.
- Alert triage: your monitoring tool calls the task's endpoint when an error threshold is crossed. The session pulls the stack trace, correlates it with recent commits and opens a draft pull request with a proposed fix.
- In-house code review: on every pull request opened, the task applies your own review checklist and leaves inline comments, so humans can focus on design.
- Documentation drift: a weekly pass flags pages that reference interfaces changed since, and opens the corrections in the documentation repository.
- Deploy verification: your delivery pipeline calls the task after each production release, which runs its checks and returns a go or no-go before the deployment window closes.
Tandem has run this blog on that principle since the summer of 2026. A scheduled task picks up the roadmap, researches its sources, produces the visuals, writes both languages and publishes to the content manager. It commits nothing and deploys nothing, by design. The guardrail that matters in editorial automation is the one separating content publishing from code deployment.
Claude Routines or n8n, what actually changes?
Both tools schedule automatic work, and the resemblance roughly ends there. n8n describes a path, node by node, and that path repeats identically on every execution. A Claude scheduled task describes an intention, and the session picks its own path on each run. The first is deterministic, the second can exercise judgement.
| Criterion | Claude scheduled task | n8n |
|---|---|---|
| What you describe | An intention in plain language | A chain of nodes |
| Behaviour run to run | Variable, the session decides | Identical, the path is fixed |
| Integrations | Connectors, plus code Claude writes | Several hundred ready-made nodes |
| Minimum interval | 1 hour | Down to the minute, or event driven |
| Hosting | Anthropic's cloud, or an enterprise environment | Cloud, your own server, or a local machine |
| Observability | A session history to reread | An execution log per node |
| Entry price | Included in the Claude subscription | Free self-hosted, around 20 euros a month on cloud |
I summed this up in a post on workflows in March 2026. On social media, nine out of ten posts announcing an agent actually show a workflow. Companies are not looking for magic, they are looking for control, reliability and return on investment. Agents make people dream, workflows keep the business running.
That reasoning breaks down into four levels of optimisation, from the conversational assistant to the autonomous agent, which I detail in my edition on agents. Most of the return on investment happens at levels two and three, simple automation and workflows enriched with artificial intelligence. A Claude scheduled task lives at level four. It is more powerful, and it is less predictable.
What it costs and which limits to know before you start
Runs are not billed separately. They draw on the usage limits of your Claude subscription, exactly like an interactive session. A chatty routine therefore eats the quota your teams use during the day, which is the real hidden cost of this feature. Organisations with usage credits turned on keep running on metered overage, the others see their runs rejected until the window resets.
On top of that sit hourly caps independent of the subscription. An account can launch up to 100 scheduled runs per hour. Each individual task accepts 30 manual launches or API fires per hour. None of these caps allows overage, the run waits or the action fails.
Comparing prices with n8n means picking a side. The self-hosted community edition stays free, and costs only the server. On the cloud side, the n8n pricing starts at around 20 euros a month for 2,500 executions on annual billing, then around 50 euros for 10,000 executions, checked in November 2026. A Claude task has no invoice line, it has a queue and a shared quota.
Cloud, desktop or session: three ways to schedule
Claude Code offers three scheduling mechanisms, and confusing them costs hours. The cloud version cannot see your local files, since it starts from a fresh clone of the repository. The desktop version runs on your machine, so it reads your files, your keys and your installed tools, provided the machine is switched on. The session loop does not survive closing the terminal.
Our own editorial automations run on the desktop version. The reason is mundane. They need the real repository, the image generation scripts and the keys living in the local environment file. A cloud task cloning the repository would have neither. Picking the right execution venue comes down to that question, ahead of any thought about how elegant the tool looks.
Letting Claude write your n8n workflows, the combination that works
Framing the two tools as rivals misses the most useful setup. Claude Code can build a complete n8n scenario from a description in plain language. You describe the flow, it looks up the right nodes, designs the architecture, creates the scenario on your instance and fixes its own formatting errors. What remains is the last mile, connecting the accounts and adjusting two or three parameters.
This combination lowers the real barrier to automation, which lies in mastering the interface, the nodes and the data mapping. The flow stays deterministic once built, since it executes inside n8n. The intelligence serves the construction, then the flow runs on its own. For a team that wants reliability, that is the best of both worlds.
Should you replace n8n with Claude scheduled tasks?
No. Not while your current flow works. Claude scheduled tasks aim at something other than an automation platform. They serve to hand a Claude Code session a piece of judgement work that comes back every week, without keeping a computer switched on. For a chain of rules across three SaaS tools, n8n stays faster to write, easier to debug and cheaper to run.
Our recommendation is therefore a clean split. Give a Claude scheduled task whatever requires reading, understanding and writing, code review, ticket triage, market watch, content production. Give n8n whatever requires executing a known path, data synchronisation, notifications, follow-ups. And have the first one write the second.
Tandem supports more than 50 companies on these calls, and the trap is always the same. Teams pick the tool before describing the task. If you are still hesitating, our n8n review gives the criteria on the platform side, and API hosted agents covers the case where neither one fits.
Two entry points to start cleanly. If the platform is new to you, begin with our n8n guide. If the project is already scoped and needs delivering, our n8n agency takes over.
