Claude Code Routines for non-developers: what to do with five scheduled agents a day

On April 14, 2026, Anthropic shipped Claude Code Routines. Nobody I know who runs a small business has read the launch post. Most of them have, at some point, complained to me about the same thing: they can’t have Claude run automatically while they’re not at the keyboard. Now they can.

A routine is a saved Claude Code configuration — one prompt, one set of connected MCP tools, one trigger — that runs on Anthropic’s servers on a schedule, through an HTTP call, or from a GitHub webhook. Your laptop can be closed. You can be on holiday. The routine still fires. If you already pay $20 for Claude Pro, you already have five of these a day. If you pay $100 for Max, you have fifteen. There’s no extra per-hour fee. The quota comes out of what you were paying anyway.

The rest of this post is how I’d actually use them, and how a non-technical person who runs an accounting firm or a dental practice should think about which ones to wire up first.

The one-paragraph version

A routine packages one prompt, one set of MCP connectors, and one trigger. Scheduled routines are cron jobs with natural-language frequencies like “every weekday at 7am” or “the first of every month.” API routines give you a private HTTPS endpoint with an auth token; anything that can make an HTTP request can fire your agent. Webhook routines currently only listen for GitHub events, which is useless for most SMBs but perfectly reasonable for any team that lives in a repo. Pro users get 5 routines a day, Max gets 15, Team and Enterprise get 25. Overage exists but Anthropic hasn’t posted pricing for it yet. The routines run on Anthropic’s infrastructure; your machine doesn’t need to be on.

That’s everything in the launch.

Why this is different from what you already had

Until April 14, to run Claude Code unattended, you had to do one of three awkward things. You could leave your laptop open all day at your desk. You could pay somebody like me to set up a Hetzner VPS with a cron job and a long-running Claude Code process. Or you could use Managed Agents, which I wrote about last week — cheaper than a VPS, but still API-only, still a job for somebody who can write code.

Routines is the first option on that list a non-developer can use without help. You open the redesigned Claude Code desktop app, you tell it what to do and when, and you save the routine. The UI is a form, not a config file. The schedule is in English. The connectors are whatever MCP servers you’ve already wired to your Claude account.

It is, finally, the cron-for-humans layer.

The math: what five routines a day actually costs

Here’s the honest cost breakdown for the three plans.

Pro is $20 a month for 5 routines a day, which works out to 150 routine runs a month. At that volume, each run “costs” you about $0.13. Most runs take a few minutes. Some take twenty. Anthropic eats the compute because it comes out of your existing Pro quota, so you burn faster against your usage limits but not against your card.

Max (5x) is $100 a month for 15 routines a day, or 450 runs a month. Each run is about $0.22. This is the plan I’d point a solo consultant or a five-person firm at. Fifteen scheduled automations is more than most SMBs will fill in six months of steady experimentation.

Team is $30 per user per month, with 25 routines a day per user. At a ten-person office where four people actually own processes, you’ve got 100 routines a day available for $120 total. That’s a lot of cron.

Compare this to the alternatives. A VPS running the same workload is $10–30 a month for the hardware, plus the hours somebody spends keeping it alive. Managed Agents costs $0.08 per active session-hour on top of your tokens, which Finout’s worked example priced at roughly $0.52–$0.70 for a moderately complex one-hour run. A daily one-hour routine on Managed Agents is north of $15 a month in session fees alone, plus tokens, plus somebody wiring it up. Routines is the lowest-friction option that exists right now. It’s the first Anthropic product where a non-developer can honestly say “I set this up myself.”

The five routines I’d wire up first for a service business

I’ve set up routines for three clients this week. These are the ones that earned their place. Pick the five that match your business and stop there. Running out of routine slots isn’t the failure mode to worry about. Building clever routines nobody uses is.

1. The 7am owner briefing

Every weekday at 7am, the routine pulls yesterday’s invoices from QuickBooks, today’s calendar from Google, unread client emails flagged as important, and any new leads from your CRM. It writes a one-page summary and emails it to you. You read it over coffee.

For an accounting firm, this is the morning you used to spend opening six tabs and checking where things stand. See my accounting industry page for the shape of the clientele. For a dental practice, it’s yesterday’s cancellations and today’s no-show risks. For property management, it’s unpaid rent reminders and overnight maintenance tickets.

Wire up Gmail MCP, Google Calendar MCP, Intuit’s QuickBooks MCP that shipped through the Anthropic partnership, and your CRM’s MCP. HubSpot and Pipedrive both have one. The prompt is one paragraph. The schedule is “every weekday at 7am.”

2. The Friday CRM sweep

Every Friday at 4pm, the routine scans your CRM for leads that haven’t been touched in more than fourteen days. For each one, it drafts a one-sentence follow-up email and saves it as a draft in Gmail. It does not send. You review Monday morning, click send on the ones you want, delete the rest.

This is the single highest-ROI automation I’ve built for service businesses in the last six months. The prompt is five lines. The value is that nothing falls through. I’ve written before about what small business owners actually want to automate first; this is usually at the top of the list.

3. Invoice chase, weekly

Every Monday at 9am, the routine checks QuickBooks for invoices more than thirty days overdue, drafts a polite-but-firm reminder, and emails it directly. For anything more than sixty days overdue, it drafts a message to you instead of the client. You decide whether the relationship can take a harder note. See my post on invoice automation for accounting firms for the longer version.

The risk here is real. An agent that emails clients without a human in the loop can embarrass you. I run these in “suggest, don’t send” mode for the first month, which means the drafts land in your Gmail drafts folder, not in the client’s inbox. The cost of over-caution is low. The cost of an over-eager agent telling a long-standing client they’re delinquent is high.

4. The content calendar fill

Every Monday at 10am, the routine reads your blog’s RSS feed, pulls the three most recent posts, and drafts three LinkedIn-length variants and three X-length variants per blog post. It saves them as drafts somewhere you can review: Notion, Google Docs, whatever you use.

For a marketing agency, this is the junior-staffer task you were paying for. For a solo consultant, it’s the thing you always meant to do and never did. The drafts will be imperfect; the point is that they exist, and editing a draft takes ten minutes instead of the hour it takes to write one.

I’ve written before about why the prompt is the product. This routine lives and dies by the quality of the prompt. A rushed one produces generic LinkedIn slop. A carefully-written one, with examples of your actual voice, produces drafts you’ll use.

5. The compliance watchlist

Once a week, the routine runs a web search for new EU AI Act guidance, new rulings from your national tax authority, or any other regulator of interest for your sector. It summarizes anything new, links the source, and emails you. If nothing is new, it sends a one-line “all quiet” message so you know the routine is still alive.

I run this for myself, Sunday evenings. The August 2, 2026 AI Act deadline is close enough now that missing a piece of guidance could matter. A five-dollar-a-month routine that watches the regulators is cheaper than the alternative.

The sixth slot

Save one for experimentation. I tried, this week, a routine that reads Monday morning Slack DMs, summarizes any question that would have reached me anyway, and drafts an FAQ entry. It worked, sort of. I turned it off. The point of the fifth-or-sixth slot is that you can afford to be wrong about one idea a week.

What Routines still can’t do, because there are always catches

Webhook routines are GitHub-only today. If your business lives in Shopify or Typeform or Calendly, you can’t trigger a routine from a webhook yet. You can approximate it by having Zapier or Make forward the webhook to an API routine, but that’s another moving part. DevOps.com’s writeup says Anthropic has hinted at expanding webhook sources; they haven’t said when.

API routines give you an HTTPS endpoint, but the auth is a bearer token with full routine-execution rights. If the token leaks, your routine runs for whoever has it. Store it the way you’d store your Stripe secret key. If you don’t know what that means, pay somebody to set it up.

Overage pricing is unspecified. The Register is blunt about this: Anthropic hasn’t said what happens when a Pro user tries to fire a sixth routine. Either it silently doesn’t run or it bills your card. You don’t want to find out by accident.

The routines run on Anthropic’s US infrastructure by default. If your workflow touches EU residents’ personal data, read Anthropic’s DPA carefully and keep sensitive data out of the routine’s prompt where you can. For sensitive workloads, self-hosted AI on a Mac Mini is still the right answer. Routines is the wrong tool for that workload.

And the obvious one: a scheduled agent that emails your customers without a human reviewer will, eventually, email the wrong thing. The “suggest, don’t send” pattern above is how I run everything client-facing for the first month. I’ve written before about why nobody notices when AI experiments fail silently; a routine that quietly sends the wrong email to fifty clients is the exact failure mode you’re trying to avoid.

How a non-developer actually sets one up

Open the redesigned Claude Code desktop app — the same one that got a full makeover the same day Routines launched. There is a Routines tab in the sidebar.

Click New Routine. Write the prompt. Pick the MCP connectors the routine will have access to. Pick a trigger (a schedule in English, an HTTP endpoint, or a GitHub webhook). Click Save. The routine is live.

That’s the setup. It is, genuinely, that simple for the scheduled case.

What isn’t simple is writing a prompt that produces the output you want. The same advice I gave in the prompt-is-the-product post applies: don’t ask Claude to do the thing directly. Ask Claude to interview you about what the thing should look like, then write a prompt from the answers. A vague routine prompt fires five times a week and wastes five slots. A specific one does real work. MindStudio’s routines guide has a decent set of starter prompts if you want templates.

Budget an hour per routine for the first few. The first one will take three. By the fifth, it takes thirty minutes.

How Routines fits with Managed Agents

They’re different products for different jobs.

Managed Agents is an API. You call it from code; it runs a session for you; you pay per active session-hour. It’s the right tool for long-running, on-demand work. A document-processing job triggered by a Zapier event. An agent that handles a few hours of customer chat at a stretch. Anything that isn’t on a fixed schedule.

Routines is a UI. You create it once; it fires on its own; the quota comes from your subscription. It’s the right tool for repeating work. Morning briefings. Weekly sweeps. Monthly reports.

Most of my clients this week are using both. The routine fires every morning. If it needs to spawn a longer-running job (say, a two-hour invoice-batch run), it starts a Managed Agents session and walks away. The session finishes on its own. You pay routine quota for the trigger, Managed Agents session fees for the long work.

For a small business with five-to-thirty staff, you probably don’t need Managed Agents at all for the first six months. Start with Routines. When a routine wants to do something that takes longer than a routine can, graduate it to a Managed Agents session.

FAQ

Do I need to be a developer?

For scheduled routines, no. You write a prompt and pick a schedule. For API routines, you need enough technical comfort to send an HTTP request with a bearer token, which is not quite “developer” but is past “never opened Terminal.”

Does the routine keep running if I cancel Pro?

No. Routines are tied to your subscription. Cancel the plan, the routines stop.

What if my routine runs for four hours?

A single routine run has a maximum duration Anthropic hasn’t fully published. In practice I’ve seen routines run up to around two hours before the system starts capping. If you need longer work, have the routine start a Managed Agents session and exit.

Can I run the same routine across multiple team members?

On Team and Enterprise plans, yes. Routines can be owned by the workspace rather than a single user, and you can share them. On Pro and Max, they’re personal.

Can I debug a routine that failed?

Yes. Every run produces an event log you can inspect from the desktop app. Same events model as Managed Agents.

What happens if the routine sends an email and a client replies?

Nothing, unless you wired a second routine to watch the inbox. Routines don’t hold state between runs except via whatever tool they wrote to (Gmail drafts, a Notion page, a file). Treat them as stateless jobs.

Is there spending control?

Less than with Managed Agents spending controls. Because quota comes out of your plan, your “cost” cap is the plan itself. That’s a feature. You can’t accidentally burn $500 in one afternoon. It’s also a limit. If your routine eats 40% of your monthly usage quota, everything else you do in Claude Code gets slower. Watch the usage meter for a week before you assume the routine is free.

Can I run routines without the desktop app?

You create them in the desktop app for now. Once created, they run on Anthropic’s servers and your machine can be off. Anthropic has hinted at a browser-based routine editor; it hasn’t shipped.

Is Routines available in the EU?

Yes. The desktop app works on both Mac and Windows. Data processing runs on Anthropic’s US infrastructure by default. An EU region is in limited preview. For anything touching EU residents’ personal data, check the Anthropic pricing page for current plan terms and your DPA coverage.

What to do this week

If you already pay for Claude Pro or Max, you have routine slots sitting unused right now. Spend one afternoon this week setting up the 7am owner briefing. You’ll know by Friday whether it’s worth the other four slots.

If you haven’t paid for Claude, and you run a five-to-thirty person service business, Claude Pro is $20. One routine, one morning briefing, will save you thirty minutes a day. That’s a fifteen-hour month for twenty dollars. Not the best ROI in the category — the invoice chase is — but the easiest to set up on day one.

If you’re not sure what to automate first, I’ve written about that too. The short version: start with the thing that currently runs through you manually every day and doesn’t need judgment. If you need help figuring out which thing that is for your specific business, that’s what I do. Pricing and scope are on the site; the first conversation is free.

The honest read on Routines today: it’s the first piece of the Claude product line where the setup is simple enough that an owner-operator can reasonably do it themselves. You still need a good prompt. You still need to know what’s worth automating. But the engineering layer, for the first time, is gone.