Prova

Prova is a tool for checking AI agent runs in production and giving a health verdict with proof attached. It says it is meant to answer whether a run worked, then provide a signed receipt that can be verified offline without trusting Prova itself. The site describes it as catching coordination loops, runaway cost, prompt injection, and PII issues in production.

Its run analysis covers every model call, agent step, and tool use as the run happens. Prova assigns each agent run a 0 to 100 health score and a letter grade, with results described as deterministic and based on signals already present in the receipts. The product also says it can flag runs, and it includes actions such as auto-rolling back a regressing deployment, blocking a disallowed call before it runs, and granting, expiring, or revoking an agent’s tools and data scopes in real time. A built-in policy library is mentioned in connection with PII, secrets, prompt injection, and data residency. The page also says it can benchmark whether a model regressed after a vendor update.

Prova is delivered as a Python SDK and is shown with a one-line install and demo command. It works with LangGraph, CrewAI, AutoGen, OpenAI Agents, Vercel AI SDK, LlamaIndex, and Pydantic AI. It also says that existing traces from LangSmith or Langfuse can be mirrored continuously into an Audit Vault without re-instrumentation. The same score can run offline in the SDK with no account, and receipts are described as Ed25519-signed and verifiable offline. The product mentions a dashboard after data ingestion, live trace analysis, and an immutable Audit Vault.

It also names a live demo and a book-a-call option.