OpenAI-compatible

The OpenAI /chat/completions format is the industry default — OpenAI, Mistral, OpenRouter, Groq, Together, Z.ai, LM Studio and Ollama (/v1) all speak it. This one zero-dependency adapter covers the whole family; vendors differ only by data (base URL, branding), which you pass as config or pick from presets.

npm install @aparte/provider-openai-compat @aparte/core

import { aparteGlobalConfig, AparteDirectTransport } from '@aparte/core';
import { createOpenAICompatProvider, presets } from '@aparte/provider-openai-compat';

// A known vendor, via a preset:
aparteGlobalConfig.registerAIProvider(createOpenAICompatProvider(presets.MISTRAL));

// …or any compat endpoint, no preset needed — just an id + base URL:
aparteGlobalConfig.registerAIProvider(createOpenAICompatProvider({
  id: 'groq',
  baseURL: 'https://api.groq.com/openai/v1',
}));

aparteGlobalConfig.setTransport(new AparteDirectTransport({ byok: true }));

Built-in presets: OPENAI, MISTRAL, ZAI, OPENROUTER, LMSTUDIO, OLLAMA.

Models are your data

Pass a static models list, or rely on the generic GET {baseURL}/models fetcher (part of the compat standard). Cloud endpoints need a key to list; local servers list keyless.

createOpenAICompatProvider({
  id: 'x',
  baseURL: 'https://x.example/v1',
  models: [{ id: 'llama-3.3-70b', name: 'Llama 3.3 70B' }],
});

Local models (LM Studio, Ollama)

Local servers are served through their OpenAI-compat /v1 endpoint. The isLocal presets relax the key requirement and fetch models keyless:

aparteGlobalConfig.registerAIProvider(createOpenAICompatProvider(presets.OLLAMA)); // http://localhost:11434/v1
aparteGlobalConfig.setTransport(new AparteDirectTransport({ byok: true }));

The full recipe — CORS, picking a model from the server’s list, llama.cpp without a preset, the context window — is Chat UI for a local LLM.

As a pure format adapter (bring your own `fetch`)

The provider separates wire format from transport, so you can take the first and keep the second. createOpenAICompatProvider returns an OpenAICompatProvider — the provider type plus the format-adapter surface, with buildRequest / parseStream / authHeaders / parseText guaranteed present (the base AparteAIProvider declares them optional, since a provider may do its own I/O):

import { readableToAsyncIterable } from '@aparte/core';
import type { AparteChatRequest } from '@aparte/core';
import { createOpenAICompatProvider, presets } from '@aparte/provider-openai-compat';

const provider = createOpenAICompatProvider(presets.MISTRAL);

const baseURL = 'https://api.mistral.ai/v1';
const apiKey = process.env['MISTRAL_API_KEY'] ?? '';
const request: AparteChatRequest = { messages: [{ role: 'user', content: 'hi' }], modelId: 'mistral-small-latest', stream: true };

// You own the call: your URL, your headers, your AbortSignal, your retries.
// `buildRequest` takes the request alone; the key goes through `authHeaders`.
const { path, body, headers } = provider.buildRequest(request);
const res = await fetch(`${baseURL}${path}`, {
  method: 'POST',
  headers: { ...headers, ...provider.authHeaders(apiKey) },
  body: JSON.stringify(body),
});

// `parseStream` returns a ReadableStream, which is NOT async-iterable in Chromium
// (or under `lib: DOM`). `readableToAsyncIterable` is core's adapter for exactly this.
const controller = new AbortController(); // yours to abort — the helper honours it
for await (const event of readableToAsyncIterable(provider.parseStream(res.body!), controller.signal)) {
  // text · thinking · tool_use · done{usage} — typed, vendor quirks already handled
  void event;
}

No transport, no AparteClient, no DOM: useful when the loop lives somewhere else entirely (a separate process, a server route, an Electron main process).

For vendors outside this family (Anthropic, Gemini…), use the AI SDK bridge.

OpenAI-compatible

Models are your data

Local models (LM Studio, Ollama)

As a pure format adapter (bring your own fetch)

As a pure format adapter (bring your own `fetch`)