Sine Gateway provides organization-scoped model access, routing, call records, and budgets. It does not require a project or a separate Gateway instance. Available models and operations depend on the environment’s published catalog and provider capabilities.
Choose managed or BYOK
| Mode | Credentials and charges | Preparation |
|---|---|---|
| Managed | The platform maintains provider credentials. Published prices are settled against the organization wallet. | Choose a published route, enable the organization’s managed budget, and provide available funds. |
| BYOK (bring your own key) | The organization saves its provider credentials. The provider bills your account; Cloud invocation service fees are currently zero. | An owner or admin creates a connection and discovers or registers models. |
Both modes enforce membership, product permissions, route restrictions, and request limits. Zero Cloud service fees for BYOK do not mean the provider is free.
Use managed models
- Sign in to Console, select an organization, and open Gateway → Settings.
- Review published routes, public Route IDs, supported operations, prices, input limits, and maximum reservations.
- An owner or admin enables managed calls and sets a UTC monthly budget and per-invocation reservation limit. New organizations start with managed calls disabled and a zero budget.
- Confirm sufficient available credits in the organization wallet. A budget limits spending; increasing it does not add funds.
- Choose a route and supported operation in Playground, provide input, and explicitly send it. Inspect results and financial status in Activity.
The platform connection appears as the read-only Sine Cloud provider. Organizations cannot edit its internal suppliers or credentials. You can enable or disable built-in platform routes for your organization, or create a single-target route to a published platform model. Multi-target routes mixing platform and BYOK models are not currently supported.
Use your own provider connection
- An owner or admin selects a provider in Providers and supplies credentials, any required Base URL, and provider-specific parameters.
- For supported discovery providers, check the connection and imported models after saving. Register other models in Models with the exact upstream model ID and supported operations.
- Copy the complete Route ID from model details. Use it as the application’s
model; do not substitute an upstream model name yourself. - Select that route and operation in Playground, submit input, and inspect the actual result.
Checking a provider’s model list does not generate content. Saving a connection does not guarantee model availability or verify credentials for providers without automatic checks. A Base URL must include the provider’s required path prefix. Custom addresses must use public HTTPS, with no credentials embedded in the URL.
Connect to the API
The Sine Gateway model API uses this public Base URL:
https://sinecloud.dev/api/gateway/v1Create an API key with Gateway access in API Keys. Model calls need modelInvoke; listing the catalog also needs modelRead. Securely load your complete key into the SINE_GATEWAY_KEY environment variable, then list routes available within its scope:
curl 'https://sinecloud.dev/api/gateway/v1/models' \
-H "Authorization: Bearer $SINE_GATEWAY_KEY"The GET /models request reads the saved catalog without generating content. Use a returned id directly as the request’s model, and check operations to invoke only operations the route declares. An empty catalog may require an administrator to configure available routes.
| Operation | Path relative to the Base URL |
|---|---|
| Chat and supported vision input | POST /chat/completions |
| Responses | POST /responses |
| Embeddings | POST /embeddings |
| Image generation | POST /images/generations |
| Asynchronous task lookup | GET /tasks/{task-id} |
A chat request supplies a real route ID, messages, and an applicable output limit. Managed requests accept max_tokens or max_completion_tokens, but not both, and cannot exceed the published limit.
Save a unique Idempotency-Key for each new generation request and retain the response’s x-sine-invocation-id. Use the current route examples in Console for media and provider-specific parameters; not every OpenAI endpoint is supported.
For optional project attribution, use x-sine-project-id. The project must belong to the same organization, and the header cannot widen a project-restricted key’s scope.
Budgets, settlement, and recovery
Token pricing settles against trustworthy actual usage. Reservations must fit the available wallet balance, remaining UTC monthly budget, and per-invocation limit. Disabling a budget does not release existing reservations. Moving into another month does not clear funds awaiting reconciliation.
Fixed-request-price synchronous operations currently require non-streaming responses and settle on a successfully delivered complete response. Asynchronous media tasks settle when the provider accepts submission; later generation failure or cancellation does not automatically refund the fee. Task queries and cancellation do not create a second model charge. Cancellation is confirmed only when the task state reports it.
Network errors, interrupted streams, and unknown usage do not prove the absence of provider charges. Inspect the original invocation or task and retain the original request identity. Reusing an idempotency key does not cause a second provider call and may return 409 with the original invocation identifier. Do not use a new key just to confirm what happened. Unknown financial status may retain a reservation for platform reconciliation.
Activity and organization Usage and Billing link related records. Overview HTTP metrics and recent records are not complete settlement statistics. Distinguish unknown usage from measured zero fees.