Declare a model served by the gateway

A gateway that has just been installed answers 404 to everything: it does not serve any model yet. This page describes the three declarations that connect it to your perimeter’s inference engine, the order in which they must exist, and what the console refuses.

The gateway connects to any engine that exposes an OpenAI-compatible API, with the open-weights model you have chosen. It is the only component that reaches the engine: neither workstations nor sandboxes know its address (security white paper V3, sections 3.4 and 8.2). The console names the engine Provider; this page says “provider” when it refers to that label.

The examples on this page declare an engine at the address https://inference.interne:8000, which serves the model qwen2.5-coder, and publish it under the endpoint code-principal.

Everything happens in the Endpoints tab, present in both build profiles. The console labels are in English; they are quoted here as you read them on screen.

The order of the three declarations

Fournisseur (Provider)  →  Modèle (Model)  →  Point d'appel (Endpoint)

A model requires the provider that exposes it, an endpoint requires the model it routes to, and the database refuses a row whose parent is missing. In the interface, the Provider selector in the model form offers no option as long as no provider is declared.

To undo, the order is reversed: an endpoint before its model, a model before its provider. The console refuses any deletion that would leave an orphan, and says so.

1. Declare the provider

Section Providers, button Add provider.

Field	Required	What it carries
Slug	yes	The engine’s short identifier, for example moteur-interne. Letters, digits, _ and - only.
Name	yes	The readable name, shown in lists.
Base URL	yes	The engine’s absolute address, scheme included, for example https://inference.interne:8000.
API Key Env	no	The name of the environment variable that carries the engine’s key, not the key itself.
API Key Header	no	The name of the HTTP header the key must be placed in.

The two key fields go together. Left empty, the gateway relays without placing a key, which suits an engine that does not ask for one. Filled in both, the gateway reads the variable from its own environment at relay time, and places its value as-is in the named header. The console stores no key: it keeps the name of the variable.

The gateway builds the address it calls by appending to the Base URL whatever the caller wrote after the endpoint name. A workstation that calls /code-principal/v1/chat/completions therefore reaches https://inference.interne:8000/v1/chat/completions. The /v1 segment is written in only one of the two places: in the workstations’ call address, as here, or at the end of the Base URL.

What the console refuses

What you entered	Response
An empty required field	400, slug, name, and base_url are required
A Slug with a space, a period, an accent	400, the accepted format is recalled in the message
A Slug already taken	409, A provider with this slug already exists
A Base URL the gateway cannot read	400, an absolute URL with its scheme is expected
A scheme other than https:// or http://	400, the gateway only relays to these two schemes

An engine reached over http://

The path from the gateway to the engine is TLS-encrypted (security white paper V3, section 3.3): declare the engine over https://. A Base URL over http:// is refused by default. The reason is written in the refusal: on that path, the requests, the responses and the engine’s key travel in clear text, and the gateway presents no client certificate to the engine; mutual authentication only applies to engines over https://.

An exception exists, outside the reference architecture: as soon as the address begins with http://, a checkbox appears in the form, and checking it counts as accepting the clear-text path, for that provider only. The tutorial Discover the gateway uses it for an engine running on the same machine as the gateway.

Two points to know:

• There is no setting that lifts this check for all providers at once. Acceptance is given provider by provider.
• The box is reset after each successful addition: it does not carry over to the next declaration.

Correct a provider address

The console exposes no way to change the Base URL, the key variable name or the header of a provider already declared. To correct an address, delete the provider and declare it again, which means deleting its models and their endpoints first.

The same applies to models. Only endpoints can be modified.

2. Declare the model

Section Models, button Add model.

Field	Required	What it carries
Slug	yes	The model name as the engine serves it.
Name	yes	The readable name.
Provider	yes	The provider that exposes it, chosen from the list.
Input price (nano€/token)	no, defaults to 0	Price of an input token, in nano-euros, whole number.
Output price (nano€/token)	no, defaults to 0	Price of an output token, same unit.

The Slug is the name the gateway writes in the model field of each relayed request, for example qwen2.5-coder. It must be the one the engine announces in its model list.

The two prices are used by the spending plans of the serverless build profile. For a model served by your own engine, leave them at 0: consumption is counted in tokens and valued at zero. The unit is the nano-euro per token, that is, a billionth of a euro.

Possible refusals: 400 if the slug, the name or the provider is missing; 409 if the slug is already taken; 400 with Provider not found if the provider has disappeared in the meantime.

3. Declare the endpoint

Section Endpoints, button Add endpoint. Two fields: Name, and Model chosen from the list. The name accepts letters, digits, _ and -; it is what clients write in their call address. A name that describes a use, such as code-principal or completion-rapide, remains valid when you replace the model served.

An endpoint is active as soon as it is created. The form offers no way to create it disabled.

Possible refusals: 400 if the name or the model is missing, 400 if the name contains anything other than the accepted characters, 409 if the name is already taken.

At this point, the gateway serves the model. The next step, pointing a workstation at this endpoint, is described in Route the IDE to the gateway.

Working with existing endpoints

Reconnect an endpoint to another model

In the Model column, click the model name: the cell becomes a selector. The change is saved the moment you choose, with no validation button and no confirmation.

This gesture changes the model served to a team without the team touching its configuration: the call address does not move, what sits behind it changes.

Enable or disable

The switch in the Enabled column writes the endpoint’s state. A disabled endpoint is refused by the gateway before any routing: the calls that reach it reach no engine. Its row stays visible in the console, greyed out.

Neither the reconnection nor the switch asks for confirmation, and a failure displays no message: the state returns to its original value when the table is refreshed. After either of these two gestures, reload the page to check that the console shows what you expect.

Delete

Deleting an endpoint has two outcomes, and the console chooses:

Situation	What happens
No call has ever been logged for it	The row is deleted.
Calls have been logged	The row is kept and disabled.

The reason for the second case: the usage history references the endpoint. Deleting it would detach that history, which attributes each call to an account and which is used, in the serverless profile, to compute spending caps. In both cases, the endpoint stops accepting calls.

The confirmation box announces both outcomes. The interface does not say which one happened: the row has disappeared, or it is greyed out.

Delete a model or a provider

Both are refused as long as a child remains:

What you are deleting	Response
A model that has endpoints	409, Cannot delete model: it still has endpoints. Delete them first.
A provider that has models	409, Cannot delete provider: it still has models. Delete them first.

There is no cascading deletion: nothing disappears unless you asked for it by name.

The required permissions

Each of these three families requires its own permission, and reading requires the same one as writing: listing the providers reveals their base addresses and the names of the variables carrying their keys, that is, the deployment topology. The exact permission names and the role that carries them are in the administration console reference.