llm-provider-registry

Give it a model name, get back where to send the request and which key to use.

"glm-4.6"            → https://open.bigmodel.cn/api/paas/v4   · ZHIPU_API_KEY    · OpenAI-compatible
"deepseek-chat"      → https://api.deepseek.com/v1            · DEEPSEEK_API_KEY · OpenAI-compatible
"kimi-k2"            → https://api.moonshot.cn/v1             · MOONSHOT_API_KEY · OpenAI-compatible
"claude-sonnet-4-6"  → https://api.anthropic.com              · ANTHROPIC_API_KEY· Anthropic SDK
"gpt-5"              → https://api.openai.com/v1              · OPENAI_API_KEY   · OpenAI SDK

A tiny, zero-dependency lookup table + resolver. It knows 16 providers — including the Chinese ones (Zhipu/GLM, DeepSeek, Moonshot/Kimi, Qwen, Doubao, MiniMax, Baichuan, Yi, Stepfun, Hunyuan) that the big frameworks make you wire up by hand.

The problem it solves

You want to let users pick any model — gpt-5, deepseek-chat, glm-4.6, qwen3-max, claude-… — and Just Work. Every one of those is reachable through the OpenAI chat-completions API, so in principle you only need one SDK. The catch is the plumbing: each provider has its own base URL and its own API-key env var, and nobody ships a maintained table of them — the OpenAI SDK obviously doesn't, and the big frameworks skip most of the Chinese providers.

So everyone ends up hand-copying base URLs off docs pages into a switch statement. This is that switch statement, curated and tested, as a 2 kB import.

import OpenAI from "openai";
import { resolveModel } from "llm-provider-registry";

function clientFor(model) {
  const p = resolveModel(model);
  if (!p) throw new Error(`Unknown model: ${model}`);
  if (!p.openaiCompatible) throw new Error(`${p.provider} needs its own SDK, not OpenAI's`);
  return new OpenAI({ baseURL: p.baseURL, apiKey: p.apiKey });
}

// One line, any provider:
const res = await clientFor("glm-4.6").chat.completions.create({
  model: "glm-4.6",
  messages: [{ role: "user", content: "hi" }],
});

Install

npm install llm-provider-registry

Node ≥ 18. ESM. Ships its own TypeScript types.

API

resolveModel(model, env?) → ResolvedModel | null

Returns null if the model isn't recognized, otherwise:

{
  provider: string;          // "Zhipu (GLM)", "DeepSeek", "Anthropic", …
  kind: "anthropic" | "openai" | "gemini" | "openai-compatible";
  baseURL: string;           // point your client here
  apiKeyEnv: string;         // e.g. "ZHIPU_API_KEY"
  apiKey: string | undefined;// that var's value, read from `env` (default process.env)
  openaiCompatible: boolean; // true → use the OpenAI SDK against baseURL
}

env defaults to process.env; pass your own object to resolve keys from somewhere else (tests, a config store, per-request overrides).

openaiCompatible is the branch that matters. It's true for everything except Anthropic (claude-*), which speaks a different wire protocol — send those to @anthropic-ai/sdk. OpenAI, Gemini (via its OpenAI-compat endpoint), and all 13 presets are true.

hasCredentials(model, env?) → boolean

true only when the model is recognized and its API-key env var is set. Handy for filtering a model menu down to the ones the user can actually call:

const usable = ALL_MODELS.filter((m) => hasCredentials(m));

PRESETS

The raw table of OpenAI-compatible third-party providers, if you'd rather iterate it yourself:

{ name: string; modelPrefixes: string[]; baseURL: string; apiKeyEnv: string }[]

Providers covered

| Provider | Example model prefixes | Base URL | API-key env | |---|---|---|---| | Anthropic | claude- | api.anthropic.com | ANTHROPIC_API_KEY | | OpenAI | gpt-, o1, o3, o4, chatgpt- | api.openai.com/v1 | OPENAI_API_KEY | | Google Gemini | gemini- | …/v1beta/openai/ | GEMINI_API_KEY | | DeepSeek | deepseek- | api.deepseek.com/v1 | DEEPSEEK_API_KEY | | Mistral AI | mistral-, codestral-, magistral-, ministral-, pixtral- | api.mistral.ai/v1 | MISTRAL_API_KEY | | Perplexity (Sonar) | sonar, sonar- | api.perplexity.ai | PERPLEXITY_API_KEY | | xAI Grok | grok- | api.x.ai/v1 | XAI_API_KEY | | Volcengine Ark (Doubao) | doubao-, ep- | ark.cn-beijing.volces.com/api/v3 | ARK_API_KEY | | Aliyun DashScope (Qwen) | qwen-, qwen2, qwen3 | dashscope.aliyuncs.com/compatible-mode/v1 | DASHSCOPE_API_KEY | | Moonshot (Kimi) | moonshot-, kimi- | api.moonshot.cn/v1 | MOONSHOT_API_KEY | | Zhipu (GLM) | glm-, chatglm- | open.bigmodel.cn/api/paas/v4 | ZHIPU_API_KEY | | Stepfun (Step) | step- | api.stepfun.com/v1 | STEPFUN_API_KEY | | 01.AI (Yi) | yi- | api.lingyiwanwu.com/v1 | LINGYI_API_KEY | | Baichuan | baichuan-, baichuan2/3/4 | api.baichuan-ai.com/v1 | BAICHUAN_API_KEY | | MiniMax | abab, minimax- | api.minimax.io/v1 | MINIMAX_API_KEY | | Tencent Hunyuan | hunyuan- | api.hunyuan.cloud.tencent.com/v1 | HUNYUAN_API_KEY |

Matching is by case-insensitive prefix, so glm-4.6, glm-4-flash, and GLM-4.6 all resolve to Zhipu — new model releases usually keep the family prefix and Just Work without a library bump.

What it deliberately isn't

• Not an API client. It hands you a baseURL + key; you bring your own SDK. That's the point — it stays 2 kB and never breaks when a provider tweaks a response field.
• Not a param translator. It doesn't reshape requests between the Anthropic and OpenAI schemas. It only tells you which family a model belongs to (kind) so you know which SDK to reach for.
• Not exhaustive on models. It matches families by prefix, not an ever-changing list of exact model IDs.

License

MIT © 2026 oratis