BYOK models

Bring-your-own-key (BYOK) models in Stratix — register OpenAI-compatible endpoints as first-class models.

Stratix supports OpenAI-compatible endpoints as custom models. Register your endpoint, store the API key encrypted in the BYOK vault, and Stratix treats it as first-class — usable in evaluations, comparisons, scorers, judges, and trace evaluations.

Why BYOK

• You're running a self-hosted model (vLLM, Together, Fireworks, Nebius, SambaNova, custom)
• You want token consumption to bill against your own provider account
• You want the same dashboard for public and private models

What "OpenAI-compatible" means

The endpoint exposes the OpenAI Chat Completions and (where supported) Embeddings APIs at a base URL. Stratix's evaluation engine sends requests in OpenAI shape and parses responses likewise.

Two billing surfaces

When evaluating with a BYOK custom model:

• Your provider bills the token consumption (what your endpoint actually does)
• Stratix ECU bills the platform work (judge calls, trace evaluation engine, etc.)

Setup

In Premium Models → Register custom model:

1. Endpoint URL
2. API key (encrypted in the BYOK vault)
3. Model identifier
4. Capability tags

Stratix probes the endpoint and confirms it responds.

Security

• Keys encrypted at rest, never returned via API (only masked prefix shown)
• Org-scoped — only members with models:write can register or rotate
• Rotation supported

Tested providers

• vLLM (self-hosted)
• Together AI
• Fireworks AI
• Nebius
• SambaNova Cloud
• MLflow gateways
• Custom endpoints behind a reverse proxy

Where to next

• BYOK custom models (Premium)
• Models
• Integrations