Quickstart

Zero to first receipt in 5 minutes.

Prova is the system of record for every AI decision your enterprise makes. Pick the path that matches what you're building.

Control plane · recommended

Instrument your agent →

Drop the SDK into a LangGraph, CrewAI, or custom run. Catch loops and runaway cost, and sign a receipt for every AI decision.

Verifier · legacy

Verify a single reasoning chain →

Paste a chain of reasoning, get a verdict on whether it has circular logic. The original Prova product, still supported.

1. Try it in 10 seconds (no account)

This hits the public demo endpoint, signs a real coordination-loop receipt, and prints a shareable URL. No signup, no key.

terminalbash

pip install prova-sdk
python -c "import prova_cp; prova_cp.demo()"
# -> public URL of a signed receipt, viewable for 7 days

2. Score your own run offline (still no account)

Run health locally: a 0-100 grade plus the signals behind it (loops, no-progress, runaway cost, PII / secrets / injection). Nothing leaves your machine. Point it at events in Prova's shape, or at raw vendor logs.

terminalbash

# events the SDK emits
prova-local --file run.ndjson

# or raw LangSmith / Langfuse / OpenAI logs, mapped offline first
prova-local --file langsmith-export.ndjson --source langsmith

3. Send signed receipts to the Audit Vault

Grab a free key at /signup (100,000 events / month, no credit card). Drop the callback handler into any LangChain or LangGraph run and every model call, tool call, and agent step becomes a signed receipt.

pythonpython

pip install "prova-sdk[langgraph]"

pythonpython

from prova_cp import ProvaClient, ProvaCallbackHandler

prova = ProvaClient(api_key="prv_...")            # your free key
handler = ProvaCallbackHandler(prova, app_id="claims-agent")

# Drop into any LangChain / LangGraph run.
graph.invoke(inputs, config={"callbacks": [handler]})

# Pass break_on_loop=True to stop the run the moment a coordination loop forms.

CrewAI, the Node SDK, and any non-LangChain runtime (via RunGuard) are in the SDK guide.

4. Verify a receipt without trusting Prova

Every receipt is Ed25519-signed. The SDK recomputes the canonical hash, fetches the public key from /api/v1/keys/{id}, and checks the signature offline. An auditor can do the same with OpenSSL.

pythonpython

from prova_cp import verify_receipt

ok = verify_receipt(receipt)   # True when the signature is valid

The full offline walkthrough is in /docs/audit.

5. CLIs

terminalbash

# Score a run offline, no account (run health + loop detection)
prova-local --file run.ndjson

# Analyze raw LangSmith / Langfuse / OpenAI logs offline
prova-local --file langsmith-export.ndjson --source langsmith

# Bulk-import existing logs into the Audit Vault (needs a key)
prova-migrate --source langsmith --file langsmith-export.ndjson

Next steps

Control plane SDK →

Python and Node clients. Auto-instrument LangGraph, CrewAI, Vercel AI SDK, OpenAI Agents, LlamaIndex, and Pydantic AI; attribute cost; verify offline.

Run health →

A label-free 0-100 verdict on every run: the score, the signals, and the pass / flag / needs-you routing.

Audit Vault + receipts →

The receipt spec and the verify-without-trusting-Prova walkthrough.

Gateway-mode enforcement →

Allow / alert / block before the model call.

Progressive delivery →

Canary a release, auto-promote on a clean verdict, auto-rollback on a signed regression.

Agent authorization →

IAM for agents: scoped, revocable, just-in-time tool and data-scope capabilities.

Fleet benchmark →

Privacy-preserving cross-customer model + regression intelligence. Contribute to see.

Transparency log →

Append-only Merkle log over your receipts. Inclusion + consistency proofs, tamper-evident even against us.

Inline detectors →

Coordination loops, prompt injection, PII, bias drift, groundedness.

Continuous ingestion →

Mirror LangSmith / Langfuse traces in continuously, no re-instrumentation.

Network discovery →

Find shadow AI from Cloudflare and Datadog logs.