What it is
A fleet budget limits how much one fleet can spend. The fleet author sets the budget inbudget under config["x-agentsfleet"] in the fleet’s Definition.
"1000000000", so the example allows 5 dollars a day and 29 dollars a month.
daily_nanos is required. monthly_nanos is optional; leave it out for no monthly limit.
Why it exists
Every workspace in one account shares the same credit balance. A fleet budget stops one fleet from draining that shared balance. The dashboard header shows what is left of that balance on every page, as a Credits chip beside the workspace switcher. It appears on screens at least 640px wide. Select the chip to open Settings → Billing and review recorded charges. The header continues to show your available balance. The fleet owns its budget. Applying a reviewed setup change replaces the limits, and deleting the fleet removes them.How it behaves
The daily limit covers a rolling 24-hour window. The monthly limit covers one Coordinated Universal Time (UTC) calendar month.agentsfleet checks the limits before a run and while the run continues. A reached limit blocks new work and stops an active run at its next billing check.
A blocked event does not retry. Raise the limit in an improvement chat, apply the change, and send a new event.
The fleet, memory, and activity stream survive a budget stop. A stopped run has no final result.
What a budget counts
A budget counts only the fleet’s own runs. The setup chat that wrote or changed the fleet is charged to the workspace owner’s balance, not to the fleet’s budget. While billing is available, a limit is checked before each run and at each renewal while it runs. A run already in progress can pass its limit by up to one renewal of usage before it stops. If billing cannot answer, work continues and can exceed that bound. The full charge is settled later.Review your spending
Open Settings → Billing to see confirmed charges across your account’s workspaces. These are called settled charges. Choose Month to date or Today, then filter by fleet or model. Both periods use Coordinated Universal Time (UTC), so a day starts at midnight UTC. The totals cover all matching activity in the selected period, including activity on later pages. Spending breakdown groups those charges by fleet and model. Each activity groups charges from one fleet run or private setup attempt. Expand an activity to see its charges and usage. Setup chat hides private draft titles and messages. Input tokens are pieces of text the model reads; output tokens are pieces it writes. The details show those counts, model names, charges and run time. When a model name is unavailable, Billing shows its recorded identifier. Some older charges combine model use and run time in one amount. Billing keeps that recorded amount because separate historical prices are unavailable. Each report includes charges confirmed by the time it was opened. Pending usage stays outside its totals; newer reports can include charges confirmed later. Missing usage information does not mean that usage was free. Changing a filter starts a new report. Moving between pages keeps the same report time, so its totals stay consistent. If loading fails, use Retry report or Retry charges. A loading error is shown as an error, rather than as zero spending. Add credits opens an email to support.Read charges through the API
GET /v1/tenants/me/billing/reports/current returns the report.
GET /v1/tenants/me/billing/activities/{activity_id}/charges returns one activity’s individual charges.
Both require billing:read and use your signed-in account.
For activity charges, include activity_type=runtime for a fleet run or activity_type=setup for a setup attempt.
To match an existing report, keep its period, fleet and model filters, and through_at cutoff.
Use starting_after to read later pages without changing those values.
agentsfleet records each provider request and each billed stretch of runtime separately.
GET /v1/tenants/me/billing/charges returns those individual records.
Each request is priced at the rate in effect when it was made, and at the tier its whole prompt falls in. Cached reads, cache writes, and output each use their own rate.
Each token category on a line is rounded up to a whole nanodollar on its own. A line can cost at most one nanodollar more per category than the exact amount, and never less.
Under a self-managed provider key, your provider bills the tokens and agentsfleet charges run time only.
Charges recorded during a billing outage settle once billing answers again.
Settlement charges the full amount and can take your balance below zero. New credit repays that debt before any remainder becomes spendable.
Limits
Zero, signs, decimal points, exponents, and numbers outside the range are refused. A JSON number in place of the string is refused too.
An invalid budget blocks review and Save with
UZ-REQ-001. A stored budget that cannot be read blocks the next run.
UZ-RUN-015 marks a budget stop during a run.
The event’s recorded failure class is budget_breach.