Budgets and quotas

A cost ceiling and a token ceiling on every run, set on the agent version. What a call cost is computed from the tokens the provider reports and the model's price, never guessed.

Every model call goes through a single place in the platform, the worker, which records for each call the tokens the provider reports, the model's price and the resulting cost. The ceilings, the Studio screens and the exports all read these records.

The ceilings of a run

A run has 2 ceilings, saved with the agent version in the Spending ceilings card of the agent's Guardrails tab:

• Cost ceiling of a run: in dollars, from 0.01 to 1,000.
• Token ceiling of a run: from 1,000 to 50 million tokens.

Both count the whole run, sub-agents included. The first ceiling reached stops the run: it ends as failed, with the exceeded ceiling as the reason. The ceilings are fixed when the run is created, so a change does not affect the runs already under way.

When the agent sets no ceiling, the project's BUDGET_MAX_LLM_COST_USD and BUDGET_MAX_LLM_TOKENS variables apply, then the default of the surface the run came from, and last a platform default of 1 USD, with no token ceiling. On the widget, the surface defaults are WIDGET_RUN_MAX_COST_USD (1 USD unless the operator changes it) and WIDGET_RUN_MAX_TOKENS (600,000 tokens).

Why a token ceiling as well

The cost of a call is computed from the model's price. A model with no price has an unknown cost, which is counted apart and adds nothing to the total, so a cost ceiling never stops it. The token ceiling does, because the provider reports tokens on every call.

A ceiling bounds what Rikoo spends on your behalf at a provider. It is a control, not a billing relationship: what you pay the provider stays between you and that provider.

Rikoo does not cap spending over a period, per project or per organization. To be warned when spending drifts, the Alerts screen has 2 rules, off by default: Cost well above a normal day (warning at 2 times, alert at 4 times the cost of a normal day) and Tokens carried per run. See Alerts.

Quotas

Plan quotas (agents, projects, members, collections, tools, MCP servers, secrets, production runs per month and more) are enforced only on the hosted offer, according to the organization's plan. A self-hosted instance is not capped by a plan; the Usage & Costs screen shows what the organization holds.

Reading the numbers

• Traces, under Monitor, shows the cost of each run, split between its own cost and the cost of its sub-agents.
• Usage & Costs shows the cost per agent over 30 days and a projection, and counts the calls of unknown cost separately instead of adding them as zero. The figures can be exported as CSV.
• With OTLP trace export configured, each model call span carries its cost (gen_ai.usage.cost) and each run span its total (rikoo.run.cost_usd). The scraped metrics carry tokens and durations, not cost. See Observability.

Models billed per unit rather than per token (a second of audio, a character, an image) are counted in their own unit, with the price from the model registry, so a bill is never doubled by adding up two different things.

The price recorded for each model, in the unit it is billed in. The metering reads these figures, it does not estimate them.