Skip to main content
Version: 3.6.1-rc.2

What is the Model Gateway?

What it is​

The platform exposes its AI features as HTTP APIs, the same ones the built-in Chat uses. They come in two kinds: the Model Gateway, for calling models, and a separate API for each RAG endpoint, for your documents. This page covers the Model Gateway.

The Model Gateway is one OpenAI-compatible entry point for every model on the platform, at https://api.<your-domain>/v1.

A client sends an API key and names a model in the request. The name can be a model ID, a model alias or, with the Responses API, a preset agent. GET /v1/models lists the chat models your key may use, and its type filter shows other kinds, such as embedding models or preset agents.

Why it exists​

Whether models are deployed on your own GPUs or provided by an external service, each would otherwise need its own URL, credentials and client setup. The gateway puts them all behind one API:

  • Switch models without changing code. Change the model value in the request, nothing else.
  • Every call is checked. Models are only reachable through the gateway, so no request skips the access check.
  • Your applications never hold provider keys. The gateway adds the provider's key itself.
  • You can see who used what. Token usage is counted per tenant, API key and model.

How it fits in the platform​

client ──► Model Gateway ──┬──► models on your own GPUs
(API key) └──► external models (OpenAI, Azure OpenAI, Anthropic, your own servers)
  1. A client sends a request that names a model, for example "model": "company-chat". The client is either:
    • your own application, which signs in with an API key, or
    • a platform feature, such as Chat when a user sends a message, or a RAG endpoint when it indexes a document.
  2. The gateway checks that the caller may use that model. What a caller can reach comes from tenancy and, for an API key, its grants.
  3. It forwards the request to the model, whether it runs on your own GPUs or at an external provider such as OpenAI, and returns the answer.

With the Responses API, the gateway can also call tools for the model, such as RAG endpoints and MCP servers, until the model has a final answer. See Tools and preset agents.

More resources​