Claude Max Usage Limits: Budgeting a 24/7 Agent Schedule

The first question people ask when I tell them my Mac runs Claude Code agents all night is not about prompts or schedulers. It's "doesn't that burn through your Claude Max usage limits?" Good question. That used to be a huge problem before models like Opus 5.5 came out that weren't both capable and efficient.

Now you can run agents on your subscription without hitting weekly limits that easily. You could run multiple animation creation agents, video creation agent, blog writing/marketing agents, and still have some left over in the morning.

It's still good to know how limits work, what an unattended agent does when it hits one, and how I manage my agents to run 24/7.

How Claude Max usage limits actually work

Pro and Max plans meter Claude Code with two clocks running at the same time:

A 5-hour session window. Short-term throughput. When it fills up, you wait for it to reset. Anthropic doubled Claude Code's 5-hour limits on May 6, 2026 and dropped peak-hour throttling for Pro and Max. It never called the doubling permanent, but the September 14 change left it alone.
A weekly cap. The long-term ceiling. It resets on a fixed schedule for your account, and it's the one that changed on September 14. A temporary 50% boost ran from May 13 through September 13. Since September 14, the weekly limit is a permanent 25% above the spring baseline.

On Max, the model families also get their own caps. Claude Code shows "You've hit your Fable limit" as a different error from "You've hit your session limit".

Two more details matter for agents. First, the pool is shared: Claude chat and Claude Code draw from the same limits, so running research agents late at night on the website could affect the limit of your scheduled agents in Claude Code.

A night shift will most hit weekly limits

While you sleep, the 5-hour window resets once or twice, so a single overnight job rarely runs into it. If one does, it can pause and pick up after the reset (more below).

The weekly cap doesn't reset overnight. Every scheduled run takes a slice of the same weekly bar you use at your desk, and since September 14 that bar is about a sixth shorter than it was all summer. Luckily Opus 5.5 (and now Sonnet 5.5) came along and people are hitting their limits way less.

What an unattended agent does when it hits the limit

This depends a lot on how the agent runs.

Interactive sessions wait and continue on their own. Since v2.1.234, when a usage limit stops Claude mid-task in an open session signed in with a claude.ai subscription, Claude Code waits and resumes the task after the reset. The docs have fine print that really matters for overnight work:

If the Mac slept for more than about 30 minutes and the limit reset while it was asleep, the session doesn't continue on its own. It waits for someone to press Enter.
It won't start waiting on its own for a reset that's more than 24 hours away. A weekly limit usually is.
After two waits in a row, it gives up with "Automatic continue stopped after repeated usage-limit hits."

The real surprise-bill trap is an API key, not the limit. If ANTHROPIC_API_KEY is set in the environment your scheduler inherits, Claude Code uses that key instead of your subscription, and every run is billed per token. Check the environment your launchd job or cron actually gets, not just your interactive shell. The same goes for usage credits. They're billed at API rates once you pass your plan limit, so if you turn them on for daytime convenience, set a monthly spend limit before you schedule anything.

Where Overboard fits

I built Overboard for Mac because I wanted this kind of budgeting to be a setting, not something I have to remember. It runs entirely on the Claude Max or Pro plan you already have. There's no API key to leak into a cron environment, and nothing leaves the Mac except the Claude session itself. What it gives you for staying inside the budget:

A concurrency cap you choose. Every scheduled slot feeds one queue, priority first and then first-in-first-out, so five jobs due at 2am don't all hit your 5-hour window at once.
Daily, weekly or every-N-hours slots, with an optional daytime window. The review can run nightly, and the dependency sweep can run weekly. A slot never overlaps its own next firing.
A keep-awake menu, so the Mac doesn't sleep through a reset and leave a session waiting for Enter.
Real sessions. Each run gets its own Herdr tab. It's a visible Claude Code session you can take over or stop, not a hidden -p pipe. The history keeps each run's closing report, so you can see what used your week.

The honest constraints: the Mac has to be logged in, and every run counts against the same plan you use by hand. It's $49 once, with a 14-day free trial. My own schedule runs blog posts, social discovery and demo-video cuts alongside the code reviews. If you're weighing this against cloud runners, I compared the two approaches in background agents: cloud VMs vs your own Mac.

Quick answers

Do Claude Code and claude.ai share usage limits?

Yes. On Pro and Max, chat and Claude Code draw from the same pool, so heavy overnight agent runs reduce what's left for chat the next day.

Can I run scheduled agents on Pro, or do I need Max?

Pro works for a couple of light jobs. Max 5x or 20x leaves room for a nightly schedule plus your working day. Measure a run in /usage first either way.

Do scheduled agents cost extra on a subscription?

Not unless you add usage credits or an ANTHROPIC_API_KEY is set in the scheduler's environment. Both switch you to per-token billing.

Did the September 14 limit change affect scheduled agents?

Yes. The temporary 50% weekly boost ended September 13 and was replaced by a permanent 25% raise over the spring baseline, a net cut of about 17% from the summer bar. The 5-hour limits stayed doubled. Re-measure any job you sized before September 14, or multiply its share by 1.2.

The takeaway

Running agents on your Claude plan swaps a bill you can't predict for a budget you can see. Treat the weekly bar as the real constraint, but let Overboard "captains" dispatch agents to do the work needed. If during the week you see usage pacing above the current allotted amount, adjust your captains to hold back usage.