Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding

Stay inside the Muse Code subscription allowance

Written October 31, 2026

For anyone paying for Muse Code by subscription. The allowance is not one number: a short window and a weekly window both count, and a prompt can be one small edit or an hour of work by several agents. The way to know what your work costs is to measure one real task, before and after, and plan from that figure.

Two windows, not one

Muse Code subscriptions are sold in tiers, and usage is metered against a five-hour window with a weekly window on top of it. The usage report shows both: the current usage window, weekly use, reset times, and your tier. In the 1.4.13 release, the muse usage command reports all four, and checking the report does not run a model turn, so reading your numbers costs nothing. Near the limit, warnings show the remaining allowance and the reset time above the message queue. That behaviour is described in the Muse Code 1.4.13 changelog summary.

The weekly window is the one that surprises people. A task that fits comfortably inside a five-hour window can still eat into the weekly allowance, and the weekly reset arrives on its own schedule, shown in the report. Meta's Muse Code product page publishes no weekly size. Do not budget from a weekly number you read somewhere else. Your own report is the only weekly figure that governs your account.

What each plan allows

Meta's product page states the plans as multiples of the cheapest plan's five-hour allowance. The cheapest plan allows 10 to 50 prompts every five hours. The range is wide because prompts are not equal: the page counts prompts, but the meter behind them counts the work each prompt sets off.

Read the plan as a multiple, not as a promise about how many tasks you will finish. Five times the allowance is five times whatever one of your tasks burns — and if your tasks are heavy, five times a small number is still a small number. That is why the measurement below comes before any decision about upgrading.

Check your allowance

Run the usage command before you start work that matters:

muse usage

Write down the four things it shows: the current usage window, weekly use, reset times, and tier. Those four numbers are your baseline. Checking them costs no model turn, so check them whenever you are unsure — before a long task, after a long task, and whenever a warning appears. For scripted checks, muse exec --json can surface the same usage state when the run has not already received a usage update, per the same changelog summary.

If your install predates the release that added this report, the command may show less than the four items above. Run muse --version and compare it with the version in the changelog summary linked above before concluding the report is empty.

Measure one real task

Pick one task you actually do — the kind you repeat most weeks, at the size you usually run it. A small bug fix, a routine refactor, a first draft of a feature. Then:

  1. Run muse usage and write down the window and weekly figures.
  2. Run the task, the way you normally would, with no changes to how you prompt.
  3. Run muse usage again and write down the same figures.
  4. Subtract. The difference is what that kind of task costs against each window, in the report's own units.

Repeat the measurement for each kind of task you run often: a fix, a review, a long build. One measurement per kind is enough to plan with. A task that burns a tenth of the five-hour window is a different planning problem from one that burns half of it, and no plan description can tell you which kind yours is. The guide to what an agent costs covers the same idea from the price side: a per-job figure from your own work, times your real job count, is the budget.

To turn a measured task into a money figure for comparison, price its tokens with the token cost calculator. That tells you what the same work would cost billed per token, which is the comparison that decides whether a subscription tier or per-token billing fits your pattern better.

Decide from the measured burn

Work out how many of your measured tasks fit in each window. If one task takes a tenth of the five-hour window, ten such tasks fill it. If it takes a tenth of the weekly window, ten fill the week. Compare those counts with the work you actually plan to do in a week, not with the work you hope to do.

Then compare plans by multiple. A plan with five times the allowance holds five times as many of your measured tasks. A plan with twenty times the allowance holds twenty times as many. If your measured week needs more tasks than your current plan holds, the next multiple up is the honest step. If your measured week fits with room to spare, a higher plan buys nothing except headroom you do not use.

When the report shows little allowance left before a reset, stop starting heavy tasks until the reset time it shows. The warning gives you both numbers — what is left, and when it comes back — so use them. Starting a long task with almost nothing left in the window means the task stalls partway through, and a stalled task costs you the work already done in it.

What can go wrong

Limits

Related playbooks

Sources

Sources: Meta Muse Code product page, Muse Code CLI: usage, Linux, and 1.4.13.

Back to: Coding playbooks hub · Also: Token cost calculator · What an agent costs · Stop an agent running up a bill · Pick the cheapest model for the job