OpenAI-compatible Provider

@karkata-ai/openai-compatible normalizes OpenAI-compatible Chat Completions requests, streams, and tool calls into the Core contract.

import { createAgent } from '@karkata-ai/openai-compatible'

const agent = createAgent({
  model: 'your-model',
  baseURL: 'https://your-provider.example/v1',
  apiKey: serverApiKey,
  maxRetries: 2,
  agent: { streaming: {}, humanInput: {} },
})

baseURL is the only required provider option. model is optional: it is sent as request metadata when provided and omitted otherwise. An application LLM proxy may ignore, override, map, or reject a client-provided model; provide it explicitly when connecting directly to a provider that requires one.

Read API keys only in a trusted server or same-origin proxy. Never place long-lived credentials in a static site, browser bundle, state, error, or message. Dynamic headers can resolve short-lived credentials per request, but a header resolver failure is not automatically retryable.

The Adapter retries network errors, 429, and 5xx only. It does not retry 401, 403, ordinary 4xx, validation failures, or invalid responses. Retries and stream reads receive the current run's AbortSignal. Use transformRequest only for vendor field compatibility and never log an unredacted request.

For manual assembly, create OpenAICompatibleAdapter and pass it to Agent as llm; both entry points use the same Runtime.

For production proxy, credential, and SSE guidance, see Production Architecture, Configuration, and the Deployment Checklist.