Helicone

Each AI call, except its content. Model, provider, tokens, cost and session. Cache use, time to first token, your properties and scores. Never the prompt or the reply.

Access

Kind: A read-only key you paste
What to give: API key with Read
Why that one: A Helicone API key with Read ticked; Helicone always adds Write, which Valistry never uses. Valistry calls only the request query endpoint, a read. Helicone sends each call's prompt and reply with it; Valistry drops them unread and keeps none.

What it attributes on its own

From the source: request; model; upstream provider; custom properties; end user (hashed); latency; what Helicone computed the request cost
Needs a rule or a trace: the provider's invoice — Helicone computes a cost, it does not bill you for inference; outcome unless a custom property carries it

The addresses it calls

https://api.helicone.ai/v1/request/query-clickhouse

Valistry calls only these addresses. It blocks any other call and logs every call.

Connecting it

Create a read key. Settings → API keys → Generate New Key Open ↗

Helicone computes a cost; the provider bills it. Connecting both Helicone and a provider behind it counts the same request twice — connect one. Pick Helicone to see cost by property or user. Pick the provider to match the invoice.