Configuration

Configuration can happen at three levels: provider discovery, SDK client construction, and an individual request.

Provider configuration

AnyLLM.create() accepts a provider name and ProviderOptions:

import { AnyLLM } from "any-llm-ts";

const llm = AnyLLM.create("openai", {
  apiKey: process.env.OPENAI_API_KEY,
  apiBase: "https://api.openai.com/v1",
  clientOptions: {
    maxRetries: 3,
    timeout: 30_000,
  },
});

The values in clientOptions are intentionally typed as Record<string, unknown> because the accepted keys differ between SDKs and versions.

Request-specific fields

Use providerOptions for request properties not represented by the common interface:

await llm.completion({
  model: "provider-model-name",
  messages: [{ role: "user", content: "Hello" }],
  providerOptions: {
    service_tier: "priority",
  },
});

providerOptions is merged into the outgoing payload after normalized fields, so it can override wire-level values. That is useful as an escape hatch, but it also means you should validate or hard-code these options rather than accepting them directly from untrusted callers.

Azure OpenAI

Azure OpenAI uses the v1 API at {endpoint}/openai/v1/. Dated apiVersion values and azureDeployment client options are rejected; pass the deployment name as model on each request. Image, transcription, and speech calls add api-version=preview unless the caller already set a query version.

export AZURE_OPENAI_ENDPOINT="https://example.openai.azure.com"
export AZURE_OPENAI_API_KEY="..."
# or
export AZURE_OPENAI_AD_TOKEN="..."

const azure = AnyLLM.create("azureopenai");

You can pass OpenAI SDK constructor fields through clientOptions, including azureADTokenProvider for Entra ID. apiKey, azureADToken, and azureADTokenProvider are mutually exclusive.

Inspect effective metadata

const metadata = AnyLLM.getProviderMetadata("openai");

console.log(metadata.envApiKey);
console.log(metadata.apiBase);
console.log(metadata.capabilities.streaming);
console.log(metadata.capabilities.pdfInput);
console.log(metadata.promptCacheKeySupport);
console.log(metadata.tier);

Metadata is returned as a clone, so changing the returned object does not modify the registry. tier mirrors the Python project's credential-backed verification matrix. It describes provider integration coverage, not whether every model and operation is available in your account.

When promptCacheKeySupport is unsupported, passing promptCacheKey to completions, Responses, or Messages fails before a network request. supported means the provider handles the key directly; passthrough means a router or custom compatible endpoint receives it unchanged.