Model Registry

The model registry tells Olyx which model identifiers a project can use. Routing, cost intelligence, trace summaries, and replay comparisons all depend on registered model metadata.

What A Model Entry Contains

Each registry entry describes one callable model path.

Field	Purpose
Name	Human-readable label in the dashboard
Identifier	Value used in the request `model` field
Provider	Provider family such as OpenAI, Anthropic, Bedrock, Gemini, Azure, or internal
Base URL	Provider or private endpoint reachable by the runtime
Visibility	Whether the model is public/provider-hosted or private/customer-managed
Region restriction	Operational boundary for the model path
Input and output cost	Token pricing used by cost intelligence
Fallback identifier	Optional alternate model used by routing policy

Register A Provider Model

Open Model Registry, select a project, and add a model.

Example:

name: GPT-4o Production
identifier: gpt-4o
provider: openai
base_url: https://api.openai.com/v1
visibility: public
region_restriction: global

Use the provider’s canonical model identifier unless you intentionally want an internal alias.

Register A Private Model

Private models are currently supported for private deployments and VPC paths. Register the model with an endpoint reachable from the customer-controlled runtime.

name: Support Classifier VPC
identifier: support-classifier-vpc
provider: internal
base_url: https://models.internal.example/v1
visibility: private
region_restriction: customer-vpc

Do not use localhost unless the runtime process that executes requests can actually reach that address.

Provider Secrets

Olyx does not require you to store provider secrets in public docs examples. For private deployment paths, keep provider credentials in the customer environment and point traffic through the configured runtime boundary.

For cloud compatibility endpoint traffic, store provider credentials according to your deployment policy and never expose them to browser clients.

Costs

Cost intelligence needs the per-token pricing fields. If a model price is unknown, set it to 0 during evaluation and update it before using cost reports for decisions.

input_cost_per_1k: 0.005
output_cost_per_1k: 0.015
currency: USD

Fallbacks

A fallback identifier should point to another model registered in the same project.

Use fallbacks for:

provider outage handling
lower-cost replay comparison
private deployment failover
progressive migration from one model family to another

Validation Checklist

Before routing production traffic, confirm:

the identifier matches the request model
the base URL is reachable by the runtime
token costs are entered
private models are marked private
the fallback identifier exists in the same project
traces show the expected selected model after execution

Next Steps

Use Routing to map intents and safety tiers to models.
Use Cost Intelligence to compare spend.
Use Traces and Replays to verify model behavior.