@anvia/openai

OpenAI provider adapter for Anvia.

Use this package when you want Anvia agents, direct completions, embeddings, image generation, speech generation, or transcription to run on OpenAI models or OpenAI-compatible endpoints.

Installation

pnpm add @anvia/openai @anvia/core

Bun

Bun 1.3.14 is the currently tested and supported runtime baseline:

bun add @anvia/openai@rc @anvia/core@rc

Compatibility tests exercise the OpenAI SDK's JSON, streaming, abort, multipart, and binary media paths without contacting a model provider.

In this monorepo, the package is available through the workspace:

pnpm --filter @anvia/openai build

Usage

import { Agent } from "@anvia/core";
import { OpenAIClient } from "@anvia/openai";

const client = new OpenAIClient({
  apiKey,
});

const model = client.completionModel({ modelId: "gpt-5.6", api: "responses" });

const agent = new Agent({
  id: "assistant",
  model: model,
  instructions: "Answer clearly and concisely.",
});

// The model-specific union is inferred: none | low | medium | high | xhigh | max.
await agent.generate({ prompt: "Solve this.", controls: { reasoningEffort: "high" } });

const result = await agent.generate({ prompt: "Summarize Anvia in one sentence." });
if (result.type === "response") console.log(result.output);

OpenAI-Compatible APIs

baseUrl changes only the endpoint. Model handles default to the Chat Completions API; opt into Responses with api: "responses" per handle:

import { OpenAIClient } from "@anvia/openai";

const client = new OpenAIClient({
  apiKey,
  baseUrl,
});

const model = client.completionModel({ modelId: "openai/gpt-5.2", api: "chat" });

The same client can create both { api: "chat" } (the default) and { api: "responses" } handles. Known reasoning models advertise a typed reasoningEffort control. For custom OpenAI-compatible model IDs, pass an explicit controls descriptor to completionModel() when the endpoint supports the same parameter. Canonical controls map to reasoning.effort for Responses and reasoning_effort for Chat Completions, taking precedence over conflicting providerOptions.

Reasoning tool-call providers

Some OpenAI-compatible chat-completions providers return reasoning in provider-specific fields while using normal tool calls. For example, Moonshot Kimi K2.6 returns reasoning_content when thinking is enabled.

The chat-completions adapter preserves this reasoning in assistant history and sends it back as reasoning_content on later turns. This matters after tool calls: providers such as Moonshot can reject the next request if an assistant tool_calls message is missing its prior reasoning_content.

For Moonshot Kimi K2.6 thinking mode:

import { generateCompletion } from "@anvia/core";

const client = new OpenAIClient({
  apiKey: process.env.OPENAI_API_KEY,
  baseUrl: "https://api.moonshot.ai/v1",
});

const model = client.completionModel({ modelId: "kimi-k2.6", api: "chat" });

const response = await generateCompletion({
  model,
  messages: chatHistory,
  tools,
  maxTokens: 16_000,
  providerOptions: {
    thinking: { type: "enabled", keep: "all" },
  },
});

Provider caveat: Moonshot rejects forced/specified tool_choice while thinking is enabled. Let the model choose tools naturally when using Kimi thinking mode.

Other Models

const embeddingModel = client.embeddingModel({ modelId: "text-embedding-3-small" });
const imageModel = client.imageGenerationModel({ modelId: "gpt-image-2" });
const speechModel = client.speechGenerationModel({ modelId: "gpt-4o-mini-tts" });
const transcriptionModel = client.transcriptionModel({ modelId: "gpt-4o-mini-transcribe" });

Use the provider-neutral helpers from @anvia/core to call media models:

import { generateImage, generateSpeech, transcribe } from "@anvia/core";

const image = await generateImage({ model: imageModel, prompt: "A launch poster." });
const speech = await generateSpeech({
  model: speechModel,
  text: "Hello from Anvia.",
  voice: "alloy",
});
const transcript = await transcribe({
  model: transcriptionModel,
  audio: { data: speech.audio.data, filename: "speech.mp3" },
});

Exports

• OpenAIClient
• structural completion, embedding, image, speech, and transcription handle types
• OpenAICompletionModelId and media model-ID types
• model constants such as GPT_IMAGE_2, GPT_4O_MINI_TTS, and GPT_TRANSCRIBE
• openai

Development

The package-local typecheck and build scripts build @anvia/core first so core subpath types are available in a fresh worktree.

pnpm --filter @anvia/openai typecheck
pnpm --filter @anvia/openai test
pnpm --filter @anvia/openai build