Skip to content

The budget, in euros

A cap in tokens says nothing about money: three million tokens on a cheap model is pocket change and on an expensive one it is an invoice. The owner decides in euros, because that is what reaches the bank account.

Providers do not bill the same way, and the system tells them apart:

Mode How it is charged How the role’s spend is calculated
Pay per use Each token is paid for tokens_entrada × tarifa_entrada + tokens_salida × tarifa_salida, per million
Fixed fee Monthly subscription (Ollama Cloud and similar) The fee split according to what each role consumed

With a fixed fee there is no price per token, and inventing one would be lying. If the plan states how many tokens it includes, the equivalent is derived (fee ÷ millions included). If it does not, the fee is split in proportion to the month’s consumption among the roles using that provider: with a single role, that role takes the whole fee.

In order:

  1. The role’s own rate, if it has been set by hand.
  2. The rate of the model assigned to it in the catalogue.
  3. If it is a fixed fee with no declared allowance, the fee split.

With none of the three, the spend is unknown, not zero. The role record says so and the euro cap is not applied: the token cap stays in place as a safety net.

Each shift records its tokens and its amount when it closes, at that day’s rate (turnos.coste, turnos.tarifa_entrada_millon, turnos.tarifa_salida_millon). A price rise three months from now does not rewrite what has already been spent.

The Agno history brings the tokens; the amount is left empty on purpose: nobody invents the price of a past shift.

  • Role record → «Budget in euros per month» and the model’s two rates.
  • Connectors → providers → billing mode, fee and included tokens.
  • Consumption (/consumo in the CLI) → spent and cap, in euros.
  • Watcher → a role that runs out of budget is the quietest stoppage of all: it does not fail, it goes mute. It surfaces as a serious alert.