Claude Scheduled Tasks: What Actually Runs While Your Computer Is Asleep
- Branden Bell
- 5 days ago
- 6 min read
A while back one of my scheduled tasks woke up thinking it was a different day than it was. It did everything right. Researched, wrote, checked its work, and got the date wrong on all of it, because the date it had been handed was stale and I had never told it to go check.
I caught it the next morning reading the post. Nothing in the run had reported a problem.
This blog runs on a scheduled task in Claude Cowork, so I have spent more time inside this feature than I planned to. The line everybody quotes is Anthropic's own, that scheduled tasks run remotely and keep their cadence even when your computer is asleep or the desktop app is closed. True. There's also a note four paragraphs down in the same article that decides whether any of that applies to your work, and I haven't seen anyone write about it.
Verified August 2026.
What a scheduled task actually is
You describe the work once and Claude keeps it. When you create a scheduled task, Claude saves your prompt as the task's instructions and runs them at the cadence you pick. Every run opens as its own Cowork session, so you can go back and read what it did like any other task.
It's not a stripped down mode either. Scheduled tasks get the same capabilities as a regular Cowork task, including your connected tools, your skills, and any plugins you've installed. Mine researches, checks the live blog, writes, builds its own cover image, publishes, then goes back and inspects the published page.
Two ways to make one, both starting from "Scheduled" in the left sidebar:
Create with Claude: you say what you want in plain language, Claude asks you what it needs, and it shows you the task name, the schedule and what it will do before you confirm.
Set up manually: you fill in the task name, the prompt, the approval mode, the frequency, and optionally the model and a working folder.
Frequency options are hourly, daily, weekly, weekdays, or manually. Scheduled tasks are available in Cowork on all paid plans, so Pro, Max, Team and Enterprise.
The sentence that decides where your task runs
All of that assumes your task can run in the cloud. Here's the note that decides whether it can:
Scheduled tasks use the built-in schedule options and work with your connectors and the files saved to your Claude account. They can't be tied to a folder on your computer.
And over in the setup steps, next to the optional working folder field, the other half:
If a scheduled task requires local files or apps, it will only run locally.
So there are two kinds of scheduled task and Anthropic never calls them two kinds. If your work lives in your connectors and in files saved to your Claude account, it runs remotely and your laptop doesn't matter. If your work needs a folder on your machine, it runs on your machine, and everything about running while your computer is asleep quietly stops applying to you.
A weekly report pulling from Slack, Drive and the web is cloud eligible. That same report is not, if it has to open a spreadsheet sitting in your Documents folder.
Worth knowing before you build, because it's usually fixable. Move the input into a connector and the task moves to the cloud along with it.
When the task has to run on your machine
A local task inherits your computer's uptime. Anthropic spells this out most plainly in the docs for scheduled tasks in Claude Code Desktop, which is a different surface than Cowork, but it's the same problem:
Tasks only run while the desktop app is running and your computer is awake. If your computer sleeps through a scheduled time, the run is skipped.
A skipped run doesn't sit in a queue waiting for you. There's a catch up mechanism and it's deliberately small. When the app starts or the machine wakes up, it checks whether each task missed any runs in the last seven days, then starts exactly one catch up run for the most recently missed time and throws away everything older. Anthropic's own example is a daily task that missed six days, which runs once on wake.
So if you were picturing six reports stacked up waiting, that's the wrong picture.
A task that wakes up late still runs
This is the one that got me, and it's a single line in the docs that's easy to skim right past:
A task scheduled for 9am might run at 11pm if your computer was asleep all day. If timing matters, add guardrails to the prompt itself.
A late run isn't a failed run. It's a complete run, doing exactly what you asked, at a time you weren't thinking about when you wrote the instructions. Mine writes and publishes, so a run confused about what day it is will publish a post dated wrong, or publish a second post on a day that already has one.
The fix has to live in the prompt because there's nowhere else to put it. Mine now shells out for the real current date in Eastern instead of trusting whatever date it's handed, and I only found that out because the two disagreed by a day on a real run. It also queries the live blog before it researches anything, and then queries it again right before it publishes, since a run that's been working for an hour can't see what another run did in the meantime.
Unattended runs stall on permission
The other quiet failure is a run that doesn't crash and doesn't finish either. If a task runs in manual approval mode and hits a tool it doesn't have permission for, the run stalls until you approve it. The session sits open in the sidebar waiting for you, which is a very different experience at 2pm than it is overnight.
And there's one category you can't clear ahead of time. Connector tools your organization has set to ask, plus MCP tools marked as requiring user interaction, prompt on every single call and don't offer an always allow option. A task that calls one of those stalls every run no matter how many approvals you've already given it.
What works is running a new task on demand once while you're sitting there, approving everything it asks for, so future runs auto approve the same tools.
What goes in every scheduled prompt I write
Writing for a task that runs while you're asleep turned out to be a different job than writing a prompt in a conversation, mostly because everything you'd normally catch in the moment has to already be on the page.
What I put in all of mine:
A guard at the top that can end the whole run: mine checks whether the work is already done and exits before spending any effort if it is.
A second check right before anything irreversible: the state at the start of a long run isn't the state at the end of it.
An instruction to verify the deliverable and not the log: a clean run report and a correct result are two separate claims, and I learned that one the expensive way.
A written definition of done, and a written definition of give up: a task with no stop condition will just keep trying.
That last one keeps showing up in everything I build. When I made a Claude skill that writes a brand kit, the piece that mattered most wasn't the writing, it was the gates that refuse to produce a file at all when the output breaks a rule, so body text under a 7 to 1 contrast ratio or a headline that falls apart at display size never ships. It stops instead of handing me something plausible.
Sources
Where I land on it
Scheduled tasks changed what I use Claude for more than any other feature, and I'd point anyone at them. The offline promise is real. It just covers a narrower slice of work than the sentence sounds like it covers.
Two questions before you build one. Does this work need anything sitting on my computer, and what does this task do if it wakes up at the wrong time. Both answers belong in the prompt.
This expands one bullet from my longer guide, how to use Claude in 2026, which covers the surfaces, the models and the plans. It's also the other half of what I wrote about what Cowork can reach when you connect a folder. If you want to see an unattended build up close, the brand kit skill is free and ungated, no email.
If you want something recurring built for your business with the guardrails already in it, that's Claude consulting with me. $250 for the first hour, then $75 an hour, hours estimated up front. The intro call is free, fifteen to twenty minutes, me and not a sales team. Whatever gets built runs in your Claude account and you own it completely.
So what would you actually put on a schedule if you trusted it to run without you?
Comments