Models and endpoints

A client is named by its client_type. An official client speaks its vendor's own API, knows the vendor's models, and reads the vendor's key from the environment. A compatible client speaks one wire protocol for any endpoint that serves it.

Use exact model ids. A model id that begins with a known family names its official client on its own, so client_type may be omitted for it; any other id raises and asks for one.

Official clients

Each row names the environment variables the client reads: the key, then the endpoint, which defaults to the vendor's own. An official client reads them whatever endpoint it is given, so it also serves a gateway that mirrors the vendor's API: OpenRouter serves deepseek/deepseek-v4-pro through deepseek-official, and SiliconFlow serves Pro/moonshotai/Kimi-K2.6 through moonshot-official, each with the gateway's base_url and api_key.

Gemini 4 Argon (gemini-4-argon) supports low, medium (default), and high thinking. MMSP maps none to low and xhigh/max to high. Its 1,048,576-token context window is shared by input, thinking, and output; synchronous streaming is limited to 65,536 output tokens. Longer outputs require Google's stored background execution, outside MMSP's stateless streaming interface. Caching is implicit only.

MiniMax-M3.1-Flash (MiniMax-M3.1-Flash) always thinks and supports low through max (default max). MMSP maps none to low and passes xhigh and max through. It has no priority tier and ignores tool_choice: "none", so fast_mode and tool_choice: "none" raise UnsupportedParameterError.

Routing by model id

Without a client_type, the family the model id begins with names the client:

The match ignores case, so MiniMax-M3 routes. A gateway's id such as moonshotai/kimi-k3, or Amazon Bedrock's global.anthropic.claude-fable-5, begins with no family and needs its client_type. The environment variable CLIENT_TYPE names one for every client the code does not.

The list in code is always current, with the base URL, the client, the modalities, the context window and the price of every entry:

import { listSupportedModels } from "@prismshadow/mmsp";

const models = listSupportedModels("CNY"); // "USD" by default

from mmsp import list_supported_models

models = list_supported_models(currency="CNY")  # "USD" by default

Compatible clients

Seven compatible clients call any endpoint that serves their protocol. Name one with client_type, and give it the endpoint and its key together: as base_url and api_key, or as OPENAI_BASE_URL and OPENAI_API_KEY in the environment (the ANTHROPIC_ pair for ant-messages, the GEMINI_ pair for google-genai, the VLLM_ pair for openai-chat-vllm-adapter, the MMSP_ pair for mmsp). A base_url other than the client's default endpoint without an api_key raises: the environment's key is not sent to an endpoint named in code (the rule). openai-chat-vllm-adapter and mmsp call it keyless instead, because their servers may take no key; without either, they reach http://127.0.0.1:8000/v1 and http://127.0.0.1:25752/v1.

Where a gateway serves more than one, prefer openai-responses. OpenRouter serves it for every model it hosts; SiliconFlow serves Chat Completions only.

const client = new AutoLLMClient({
  model: "qwen/qwen3.8-27b",
  clientType: "openai-responses",
  baseUrl: "https://openrouter.ai/api/v1",
  apiKey: process.env.OPENROUTER_API_KEY,
});

client = AutoLLMClient(
    model="qwen/qwen3.8-27b",
    client_type="openai-responses",
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ["OPENROUTER_API_KEY"],
)

Common base URLs, and the variable the examples read each key from:

TokenDance serves several protocols with one key. Its model list names the protocols of each model; pick a client that speaks one of them, such as deepseek-v4.1-flash through openai-chat, and the vendor's official client where TokenDance speaks the vendor's API: kimi-k3 through moonshot-official, glm-5.3 through zai-official.

OpenCode Go also serves several protocols with one key, each at its own base URL, and its docs name the protocol of each model: openai-responses for gpt-6-luna, gpt-5.6-luna, grok-4.7 and the Muse Spark models; ant-messages at the Anthropic base for claude-haiku-5-5, minimax-m3 and the Qwen models; google-genai at the Google base for gemini-4-argon; and openai-chat for the rest. On Fireworks AI, DashScope, Qwen Token Plan and Atlas Cloud every model takes openai-chat, and on ModelScope openai-responses. Penguin Go is an MMSP server, so its models take mmsp; a vLLM server's take openai-chat-vllm-adapter.

Gemini on Vertex AI

Name google-genai and pass the service-account JSON key as the API key. Vertex AI's Interactions endpoint serves none of these models, so google-official, which speaks Interactions, refuses such a key. The client calls the global endpoint, https://aiplatform.googleapis.com, with or without that base_url.

const client = new AutoLLMClient({
  model: "gemini-3.8-flash",
  clientType: "google-genai",
  apiKey: fs.readFileSync("service-account.json", "utf8"),
});

client = AutoLLMClient(
    model="gemini-3.8-flash",
    client_type="google-genai",
    api_key=open("service-account.json").read(),
)

Thought signatures replay only on the endpoint that issued them. See Fidelity.

Claude on Amazon Bedrock

Name anthropic-official, pass bedrock://<region> as the base URL and the AWS access key and secret key, joined by a comma, as the API key. The model id is a Bedrock inference profile such as global.anthropic.claude-opus-5-5; the global. profiles cost the Anthropic list price, the geo profiles (us., eu., jp., au.) 10% more.

const client = new AutoLLMClient({
  model: "global.anthropic.claude-opus-5-5",
  clientType: "anthropic-official",
  baseUrl: "bedrock://us-east-1",
  apiKey: process.env.BEDROCK_API_KEY, // "ACCESS_KEY,SECRET_KEY"
});

client = AutoLLMClient(
    model="global.anthropic.claude-opus-5-5",
    client_type="anthropic-official",
    base_url="bedrock://us-east-1",
    api_key=os.environ["BEDROCK_API_KEY"],  # "ACCESS_KEY,SECRET_KEY"
)

Wire protocols

Every client speaks one protocol on the wire, whichever client_type reaches it.