Spend

Catch the cost. Cap the run.

Every model call carries a signed USD cost. Set a monthly cap and Prova blocks at the gateway before the call executes. Per-app, per-agent attribution from one source of truth: the receipt.

What it is

Two pieces wired into the control plane:

  • Cost on every receipt. The SDK extracts token counts from OpenAI, Anthropic, Bedrock, and Gemini responses. The server computes USD from a maintained pricing catalog and signs the cost into the receipt's integrity block.
  • monthly_budget_cap policy. 17th built-in policy. Reads a per-org counter incremented atomically by an audit_events trigger. Returns action=block on the gateway before the call runs.
  • Spend dashboard. Today, projected end-of-month, top spenders by app, cap-progress bar that turns red at the alert threshold.
  • Webhook on block. Same dispatch path as policy / detector findings. Wire it to Slack or PagerDuty.

What it does for you

Stop the AI bill from being a surprise.

The dashboard shows today's spend, projected end-of-month, and the cap-progress bar in real time. Devs see the bill grow as their agents run. Engineering leads see it before finance does.

Hard cap, not soft warning.

Portkey and Helicone observe. Prova enforces. When monthly spend crosses your cap, the gateway returns action=block with a signed receipt of the denial. Your agent runtime sees a 200 with a refusal, not a $10k overrun.

Per-agent attribution, not per-key.

Most cost tools aggregate by API key. In a multi-agent system that hides which agent is burning the budget. Prova ties cost to source.app_id and source.framework, so the dashboard says 'the reviewer agent in claims-orchestrator is your top spender,' not 'key prv_abc... used $4k.'

The signed receipt is the moat.

Cost lives inside the Ed25519-signed integrity block. The number Finance sees, the number the auditor reconciles, and the number the dashboard displays are the same number, derivable offline from the receipt and a public pricing snapshot. Helicone can copy the chart. Nobody can copy the signed audit trail.

How it ships

  1. pip install prova-sdk. Drop ProvaCallbackHandler into your LangGraph or CrewAI runtime. Token counts and cost start flowing on the next invocation.
  2. Visit /dashboard/spend. Watch today's spend climb in real time.
  3. Set a cap. The monthly_budget_cap policy takes effect on the next gateway-check call.
  4. Wire a webhook for the block event. Now Slack pings the team before the AI bill turns into a postmortem.
START ATTRIBUTING COST →Read the SDK docs