External models

An external model is a model another provider serves, called through the Olympusx AI gateway like any catalog model. Tenants call it with their own Olympusx AI keys; grants, budgets, guardrails and metering apply to it as to a model you host. Nothing is deployed for it. Platform admins register external models on Build › Models, on the External tab.

Who can open it

Role	What they can do
Platform admin	Holds every capability: register, test, edit and remove external models, and manage their connections
Admin	See external models in the catalog, marked Served by provider, and use them once the tenant holds a grant
Builder	See external models in the catalog and choose them for agents and assistants
Auditor	See external models in the catalog
Reviewer	No access
End user	No access

The External tab appears only for platform admins.

Providers

Provider	Notes
OpenAI	Default endpoint and data residency filled in for you
Anthropic	Claude models, through Anthropic’s OpenAI-compatible API
Oracle OCI Generative AI	Needs a region and a compartment (or tenancy) OCID. Authenticates with an OCI Generative AI API key, or with workload identity where the platform’s cluster supports it, which stores no secret
OpenAI-compatible endpoint	Any endpoint that speaks the OpenAI API, such as Azure OpenAI, Google Gemini (base URL `https://generativelanguage.googleapis.com`, path prefix `/v1beta/openai`), or a model server you run elsewhere

Oracle OCI Generative AI can also be reached as an OpenAI-compatible endpoint with an API key and the path prefix /openai/v1. When the base URL you type is Oracle’s, the wizard says so and offers Use /openai/v1. On that API a model on a dedicated AI cluster is called by its endpoint OCID: type it as the model id. The key needs an OCI IAM policy that lets it use Generative AI in the compartment.

What you see

The tab opens with a line on what an external model is, then a table with a search box (Search external models) and two buttons: Connections, which opens Settings › Model providers, and Register external model.

Column	What it holds
Model	Display name, linked to the model’s page, with the catalog name beneath
Provider	The connection’s name and kind, its data residency, and Connection disabled or Connection failing when either applies
Type	LLM or Embedding
Tool calling	Can call tools, Cannot call tools or Not known. A dash for an embedding model
Context window	In tokens, or a dash when nobody has said
LLMaaS	Offered when tenants can subscribe to it
Last test	Answered, Did not answer with the provider’s reason, Not tested, or Testing…
Actions	Test, and a menu with Edit capabilities…, LLMaaS offering…, Manage connection and Remove from catalog

With no external model registered, the tab shows No external models yet and Register external model. If the catalog cannot be read, it shows The model catalog could not be loaded with a retry.

What you can do

Register an external model

Go to Build › Models and open the External tab.
Select Register external model. The wizard has three steps.
Connection. Choose New connection or Existing connection (n).
- For an existing one, pick it in Connection. Only enabled connections are listed.
- For a new one, choose the Provider, give it a Name, the Base URL where it is required, the region and compartment for Oracle, the API key, and the Data residency. Advanced holds the Path prefix (optional) and the Credential header (optional).
Select Next. A new connection is created and tested. If it does not answer, it is removed again and the step says why. If the endpoint already has a connection, the wizard offers Use name instead.
Models. Pick models from Models the provider lists, or type a Model id exactly as the provider names it (such as gpt-4.1-mini or claude-sonnet-4-5), choose its Type, and select Add. Many endpoints list nothing, or not everything: when an endpoint lists no models, the step says so and you type the ids.
For each model under To register (n), set:
- Display name (optional): what people see in pickers.
- Tool calling (LLMs only): Can call tools, Cannot call tools or Not known.
- Context window (optional): in tokens, as the provider documents it.
Select Next.
Register. Review the Catalog names. Leave Send each model one short test request on to prove each model id, and Offer as LLMaaS on to let tenants subscribe to the models that answer.
Select Register.

The result reports Registered n external model(s). A model that fails its test request is registered but not offered; its row says why. Fix the id or the connection, then test it again.

The test request costs the platform a few tokens at the provider. It proves the model id, which the connection test cannot: a mistyped id authenticates and then fails every request.

Catalog names

Each model is named connection:model id in the catalog, lowercased, with any / in the id turned into -. A model gpt-4o on a connection named openai becomes openai:gpt-4o. That is the name tenants call. Name connections with letters, digits, ., _ and - only.

Test a model

Select Test on its row. Olympusx AI sends one short request through the same path tenant traffic takes: a chat request that asks for one word, or an embedding of one word. The toast reports The model answered with the time it took, or The model did not answer with the provider’s reason. Last test keeps the result.

Edit capabilities

Select Edit capabilities… on the row.
Change the Display name, Tool calling or Context window (optional).
Select Save.

Requests and prices keep using the catalog name. If agents on the model use tools and you set Cannot call tools, the dialog counts them: they will fail whenever they call a tool, skill or collaborator.

Offer a model to tenants

Select LLMaaS offering… on the row, set Offered as LLMaaS to Yes, and select Save. An external model uses none of the platform’s GPUs, so each of its tokens counts as one against a budget.
Price it in Billing. See Billing.
Grant it to tenants: open the tenant on Customers › Tenants and set its subscription and budgets. See Tenants.

Grant an external embedding model to a tenant before that tenant’s knowledge uses it. A collection on an embedding model the tenant holds no grant on is refused, so documents never reach a provider without that decision.

Manage the connection

Manage connection opens Settings › Model providers, where you rotate a key, edit or disable a connection, or delete it. A connection that still serves catalog models cannot be deleted: remove its models first. See Settings.

Remove an external model

Select Remove from catalog on the row.
Confirm.

Tenants lose access to it and their requests to it fail. The connection stays, and the model can be registered again.

Provider health

System health has a Model providers group: one tile per enabled connection, checked by listing its models, which costs no tokens. On this tab, a connection that last failed its test shows Connection failing under each of its models. See System health.

Limits

Rule	Value
Types an external model can have	LLM, Embedding
Context window	A whole number of tokens, 1 to 100,000,000, or empty
Test request timeout	60 seconds
Endpoints the gateway forwards to a provider	Chat completions, completions, embeddings
API key	Stored encrypted and never shown again
Connection kind	Cannot change once added; add another connection instead