Skip to main content

For Testing Platforms

You run a testing and evaluation platform and you want your customers to connect their private AI systems to it. These systems sit behind firewalls you can't reach, and every new customer is a different network, a different security team, a different set of constraints. You need a single protocol that works across all of them.

What you run​

You run a relay server: a WebSocket server that accepts connections from spectral-bridge clients and forwards requests from your platform's internal services. You implement it according to the protocol specification.

The relay server exposes two endpoints:

  • WebSocket /connect (external): customers' spectral-bridge clients connect here over WebSocket, authenticating with a Bearer token. This is the only endpoint exposed to the outside.
  • POST /v1/chat/completions (internal): your platform's own services call this to send requests to a connected client. This endpoint should not be publicly reachable.

When your services send a request to the internal endpoint, the relay server wraps it in a frame and pushes it down the WebSocket to the connected client. The client's response travels back the same path.

What you give the customer​

To connect, a customer needs three things from you:

  • Setup instructions: how to install the spectral-bridge client and configure it with an adapter. See For AI Teams for the customer-facing perspective.
  • A relay URL: the WebSocket address of your relay server (e.g., wss://relay.example.com/connect).
  • An API key: a token that authenticates the client on GET /connect.