Gateway/Get started/Migrate from OpenRouter

Migrate from OpenRouter

Model slugs match OpenRouter. Swap the base URL and the key; the rest of your code stays.

Change two lines

const client = new OpenAI({
  // baseURL: "https://openrouter.ai/api/v1",
  baseURL: "https://api.inferencebureau.com/v1",
  apiKey: process.env.IB_API_KEY,
});

What stays the same

Model slugs, such as openai/gpt-4o-mini.
The chat completions request and response shape, including streaming, tools, and response_format.
Token usage fields in the response.

What is different

	Inference Bureau
AI SDK	Replace `@openrouter/ai-sdk-provider` with `@inferencebureau/ai-sdk-provider`: `ib("openai/gpt-4o-mini")` takes the same model slugs. See AI SDK.
Model fallback	Off. The model you request is the model that answers, unless model fallback is allowed in Directives. OpenRouter's `models` list is accepted and ignored.
Provider choice	Dispatch tries certified providers in a fixed order (see Routing). Limit them with `allowedProviders` or `deniedProviders` in Directives, not per request.
Presets	OpenRouter's `@preset/…` models become team aliases: send `org/fast` as the model, and define it once in Directives as a model id or a preset (model, fallback models, and defaults for `temperature`, `top_p`, max output tokens and reasoning effort). Values in the request win over the preset's.
data_collection	Instead of `provider.data_collection: "deny"` per request, set `data.requireNoTraining` in Directives to use only providers listed as not training on prompts; `requireZdr` requires zero data retention. Per-request `provider` preferences are not used.
max_tokens	Without `max_tokens`, output is limited to 4,096 tokens, because every request reserves its worst-case cost up front. Set it explicitly for longer answers.
Balance and limits	Prepaid: requests are refused with `402` when the balance can't cover them. Every team has default rate limits (600 requests and 2,000,000 tokens a minute, 10 at once); see Rate limits.
Costs	Look up any request with `GET /v1/requests/:id` for its final charge.
Headers	OpenRouter-specific headers such as `HTTP-Referer` and `X-Title` are not used. Send `X-Client-Request-ID` for your own correlation id.
Claude models	Chat completions works, but tools and `tool_choice` go through an OpenAI-shaped translation. For Claude tool use, especially forced tools, call `POST /v1/messages` with the Anthropic SDK: requests pass through natively.
Web search	OpenRouter's `:online` suffix and `plugins` aren't supported. Claude searches through the `web_search` tool, and Perplexity Sonar searches on every request. See Web search.
Retries	If a provider fails before it starts answering, the request is retried on the same model, at most three attempts in all. Each attempt is listed on the receipt. See Routing.
End users	The `user` field works as on OpenRouter: it is recorded as your end user and never forwarded. Logs, usage export and per-end-user limits and budgets in Directives use it. See End users and tags.
Tags	Send `ib-tags` (or `ib.tags`) to break costs down by feature. They are kept by Inference Bureau and cost nothing extra.
Embeddings	`POST /v1/embeddings` works for certified embedding models; `GET /v1/models` lists them with `type: "embedding"`.
Responses API	`POST /v1/responses` works on models certified for it. Responses are kept only with `store: true` (OpenRouter keeps none), for 30 days, so `previous_response_id` can continue them. OpenAI's `web_search` needs `max_tool_calls`; its other hosted tools are refused.

Only certified models can be called. GET /v1/models lists the ones your key may use, with their capabilities, and Models shows the full catalog with certification status.