Custom adapters

Create a dedicated adapter when a provider has a different wire format, authentication scheme, or response model. Extend BaseProvider, implement metadata and completion(), and override any other supported operations.

import {
  BaseProvider,
  registerProvider,
  type ChatCompletion,
  type ChatCompletionChunk,
  type CompletionParams,
  type ProviderMetadata,
  type ProviderOptions,
} from "any-llm-ts";

const metadata: ProviderMetadata = {
  name: "company",
  documentationUrl: "https://developers.example.com/llm",
  requiresApiKey: true,
  envApiKey: "COMPANY_LLM_API_KEY",
  capabilities: {
    completion: true,
    streaming: true,
    embedding: false,
    listModels: false,
    responses: false,
    vision: false,
    reasoning: false,
    moderation: false,
    imageGeneration: false,
    audioSpeech: false,
    audioTranscription: false,
    batch: false,
    messages: false,
    rerank: false,
  },
};

class CompanyProvider extends BaseProvider {
  readonly metadata = metadata;

  constructor(private readonly options: ProviderOptions = {}) {
    super();
  }

  async completion(
    params: CompletionParams,
  ): Promise<ChatCompletion | AsyncIterable<ChatCompletionChunk>> {
    // Call the provider and normalize its response here.
    throw new Error("Implement company completion");
  }
}

registerProvider(
  "company",
  (options) => new CompanyProvider(options),
  { metadata },
);

After registration:

const llm = AnyLLM.create("company");

Adapter responsibilities

validate required credentials and endpoint configuration
translate common request fields to the provider protocol
normalize non-streamed and streamed responses
preserve unmodeled data in raw or extraContent
normalize SDK errors, including errors raised during stream consumption
reject unsupported operations clearly

Use the protected execute() and protectStream() helpers from BaseProvider to apply common error normalization around provider calls.

Adapter responsibilities

On this page