OpenAI compatibility

OpenToken speaks the OpenAI Chat Completions wire format. If your code already talks to OpenAI, you point the OpenAI SDK at OpenToken by changing two things — the base URL and the API key — and keep the rest of your call sites unchanged.

Swap the base URL and key

Set the base URL to https://api.opentoken.kr/v1 and pass an OpenToken key (these start with sk-optk-).

from openai import OpenAI

client = OpenAI(
    base_url="https://api.opentoken.kr/v1",
    api_key="$OPENTOKEN_API_KEY",
)

resp = client.chat.completions.create(
    model="google/gemini-3-flash",
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.opentoken.kr/v1",
  apiKey: process.env.OPENTOKEN_API_KEY,
});

const resp = await client.chat.completions.create({
  model: "google/gemini-3-flash",
  messages: [{ role: "user", content: "Hello" }],
});
console.log(resp.choices[0].message.content);

What works

• POST /v1/chat/completions with the OpenAI shape — system, user, assistant, and tool message roles.
• Non-streaming responses (chat.completion) and streaming with stream: true, which returns Server-Sent Events terminated by data: [DONE].
• The OpenAI sampling parameters: temperature, top_p, top_k, max_tokens (alias max_completion_tokens), stop, presence_penalty, frequency_penalty, seed, and more. Support is per-model — each adapter forwards only the params its upstream accepts and silently drops the rest (e.g. Claude Opus 4.7+ drops temperature/top_p/top_k; Gemini 3.x drops presence_penalty/frequency_penalty/seed). See /docs/guides/sampling-parameters.
• tools / tool calling are supported on Google (google/*) models — translated to the provider's native function calling — but rejected by Anthropic models, which return an unsupported_parameter (400) error; use response_format for structured output there. response_format (json_object or json_schema) is supported across providers (Anthropic maps json_schema to constrained decoding; json_object is best-effort).
• GET /v1/models for listing the registered models, and POST /v1/embeddings for the google/text-embedding-004 embeddings model.
• The official OpenAI SDK in any language, plus any OpenAI-compatible HTTP client.

What differs

• Model ids are {provider}/{model}. Use the namespaced id (for example google/gemini-2.5-pro), not a bare OpenAI model name. See /docs/models.
• Google and Anthropic models are registered today — for example google/gemini-2.5-pro, google/gemini-3-flash, google/gemini-3.1-pro, google/gemini-3.1-flash-lite, anthropic/claude-opus-4-7, anthropic/claude-sonnet-4-5, and anthropic/claude-haiku-4-5. OpenAI (openai/*) ids are not registered and return a model_not_found error. See /docs/models for the full list.
• Embeddings are supported via POST /v1/embeddings for the google/text-embedding-004 model. Calling /v1/embeddings with a model whose provider does not support embeddings returns an embeddings_unsupported (400) error.
• Some superset parameters are silently dropped per provider. OpenToken accepts the OpenRouter sampling superset, but each adapter forwards only what its upstream supports. For example, google/gemini-2.5-pro accepts frequency_penalty, presence_penalty, and seed, while google/gemini-3.1-pro rejects them, so they are dropped.

Next steps