Get started
Key concepts
The handful of ideas everything else is built from.
ProviderWhere models run: OpenAI, Anthropic, OpenRouter, Azure, AWS Bedrock or your own servers. Each is marked on-prem, private cloud or public, and records the contracts you hold (BAA, DPA, SOC 2…).ModelA specific model at a provider. Models are imported automatically and typed (chat, embedding, image, speech, rerank, moderation), so each endpoint only uses the right kind.Governance profileThe rule bundle an app follows: frameworks, what happens to each kind of sensitive data, safety checks, what's stored and for how long. Each API key has exactly one.FrameworkA regulation or standard such as HIPAA, GDPR, the EU AI Act or SOC 2. Choosing one on a profile turns it into enforced routing limits. When several apply, the strictest wins.Policy ruleDecides which models may answer, for example "personal data never goes to public APIs". Rules are checked in order; the first match decides; nothing matching means the request is refused.API keyWhat an app uses instead of a vendor's key. It carries the app, team and profile, so the protection depends on which key is used. Keys can expire and be revoked instantly.Workload (app)The registered AI use behind a key: its purpose, owner, data, EU AI Act risk tier and budget. Only running workloads can call models.RedactionReplacing sensitive values with placeholders such as [REDACTED:EMAIL] before the text leaves. Usually applied only when data leaves your data centre.Safety modelA small model, such as Llama Guard, that checks requests and answers for unsafe content. "Automatic" uses one on your own providers, never a public one.Audit logEvery decision and configuration change, chained with hashes so any tampering is detectable. You can verify the chain and export it.
How they fit together
An app calls the gateway with its API key. The key points to a governance profile, which applies its frameworks and data handling. Policy rules then pick from the models your providers offer, and everything lands in the audit log.