API overview

NoviaHub is compatible with three API protocols: OpenAI, Anthropic and Google Gemini. You can use the official SDKs from all three: point them at NoviaHub and use the API key you created on NoviaHub.

Endpoints

Protocol	Endpoint	Method and path
OpenAI	Chat Completions	POST /v1/chat/completions
OpenAI	Responses	POST /v1/responses
OpenAI	Image generation	POST /v1/images/generations
OpenAI	List models	GET /v1/models
Anthropic	Messages	POST /v1/messages
Anthropic	List models	GET /v1/models
Gemini	generateContent / streamGenerateContent	POST /v1beta/models/{model}:generateContent
Gemini	List models	GET /v1beta/models

The full URL is https://noviahub.com plus the path, for example https://noviahub.com/v1/chat/completions.

Which endpoints a model supports is shown by its endpoint labels on the Models & Pricing page. See Protocol conversion.

Setting the SDK base URL

Each SDK appends its own path to the base URL, so the three are written differently:

SDK	Base URL	Note
Official OpenAI SDKs (Python, Node.js, …)	https://noviahub.com/v1	Include /v1
Official Anthropic SDKs	https://noviahub.com	Do not include /v1; the SDK adds /v1/messages
Google Gen AI SDKs	https://noviahub.com	Do not include /v1beta; the SDK adds /v1beta/models/...

Authentication: where the API key goes

Every request must carry your API key. NoviaHub accepts the following, but each form only works on certain paths:

Form	Where it works
Header Authorization: Bearer <key>	Every endpoint (recommended)
Header x-api-key: <key>	Only paths containing /v1/messages or /v1/models
Header x-goog-api-key: <key>	Only paths starting with /v1/models, /v1beta/models or /v1beta/openai/models
Query parameter ?key=<key>	Same as above: only paths starting with /v1/models, /v1beta/models or /v1beta/openai/models

Details we tested:

• The sk- prefix is optional: both Bearer sk-abc... and Bearer abc... work.
• Leaving out Bearer and putting the key straight into Authorization also works.
• Using x-api-key or x-goog-api-key on a path that doesn’t accept it (for example x-api-key on /v1/chat/completions) counts as sending no key and returns 401.
• Don’t append -<number> to the key (such as sk-abc...-1). For regular users this returns 403 with the Chinese message 「普通用户不支持指定渠道」 (“regular users cannot choose a channel”).

Request ID

Every response carries the header X-Oneapi-Request-Id with the ID of the request. Error messages also end with (request id: ...). When you report a problem, please include this ID so we can trace it. It is also shown as Request ID when you open an entry in Usage Logs in the console.

Error format

Errors come with a non-200 HTTP status and a JSON body in one of two shapes.

OpenAI shape: used for every error from the OpenAI and Gemini endpoints, and for errors raised by any endpoint during authentication, model selection and rate limiting (including /v1/messages).

Example: missing messages field (measured on a test instance)

{
  "error": {
    "message": "field messages is required (request id: 2026...)",
    "type": "new_api_error",
    "param": "",
    "code": "invalid_request"
  }
}

Anthropic shape: used by /v1/messages for errors raised after authentication and model selection.

Example: calling /v1/messages with no balance left (measured on a test instance)

{
  "error": {
    "type": "new_api_error",
    "message": "用户额度不足, 剩余额度: ＄0.000000 (request id: 2026...)"
  },
  "type": "error"
}

• If the error comes from the upstream model service, NoviaHub passes through its HTTP status and error body, and also its Retry-After header when there is one.
• The Accept-Language header picks the language of error messages. With Accept-Language: zh-CN, an invalid key gives 「无效的令牌」; without it, Invalid token. Some messages currently exist in Chinese only, such as insufficient balance (「用户额度不足」) and IP not allowed.

For what each error means and how to fix it, see Errors and troubleshooting.

Other limits

• Request body size: the gateway caps the size of a request body. The built-in default is 128 MB; larger bodies get 413.
• Streaming idle timeout: during a streamed response, if the upstream sends no data for a while, the gateway ends the stream. The built-in default is 300 seconds.

Both numbers are built-in defaults that the platform may change; the actual response is what counts.

Unsupported endpoints

NoviaHub does not currently offer the endpoints below; calls return an error right away (measured on a test instance):

Endpoint	What it returns
POST /v1/messages/count_tokens (Anthropic token counting)	404, Invalid URL (POST /v1/messages/count_tokens)
POST /v1beta/models/{model}:countTokens (Gemini token counting)	404, Invalid URL (...)
/v1/files, /v1/fine-tunes, /v1/images/variations	400, Model name not specified, model name cannot be empty
Any other unlisted /v1/... path	404, Invalid URL (METHOD path)

NoviaHub also has no embedding, rerank, speech or video models yet. Calls to those endpoints return 503 because no channel is available.

Don’t use GET /v1/models/{model} to check a model

This single-model lookup only searches a built-in list. When the model isn’t on it, it returns HTTP 200 with a model_not_found error. In our tests it reported models such as claude-sonnet-5, which work fine on NoviaHub, as “does not exist”. To see which models your key can use, call GET /v1/models.