ControlWhat it doesDefault
Request logging & pricingEvery request — including failures — is logged and priced against the versioned rate table.Always on
Waste scoreA 0–100 score with named flags on every completed request: repeated prompt, large context, premium model on a small task, low output vs input, latency, error waste.Always on
Provider credential encryptionAES-256-GCM with a master key held outside the database. Only a masked hint is ever displayed.Always on
Exact response cacheDeterministic requests only — temperature 0, no tools, no response format. 24-hour entries; a hit costs nothing and records the original call's cost as the saving.Off → Recommend → Auto
Model routingSmall, simple, low-temperature GPT-4-family calls can move to gpt-4o-mini. Features or endpoints named legal, medical, financial or security are never downgraded.Off → Recommend → Safe auto
Prompt-injection scan (inbound)Scores untrusted input; enforce refuses high-risk requests with 403 before a provider is called.Off → Detect → Enforce
Secret-leak scan (outbound)Signature and entropy detection on the response; enforce replaces the value with a redaction marker.Off → Detect → Enforce
Personal-data scan (outbound)Emails, cards, SSNs, phones, IPs, IBANs; enforce masks them before the response is returned.Off → Detect → Enforce
Prompt storageRequest bodies are stored only when you enable it per project — for richer recommendations. Responses are never written to the request log; metadata and one-way hashes are always kept.Off by default
Provider failoverOne retry, then your secondary provider — for transient failures only (5xx, 429, network). A 4xx is returned as-is.Configurable
Rate, token & spend ceilingsRequests per minute, maximum prompt and output tokens, and a monthly spend cap enforced before the provider is called. New projects start at 60/min, 8,000, 1,024 and $5.Configurable
Streaming responsesstream: true forwards SSE chunks in the OpenAI format across every provider. Cost, tokens and the waste score are computed on the assembled response when the stream closes.Supported (except under outbound enforce)