llm-provider-catalog

GitHub

A community-maintained registry of LLM provider metadata — base URLs, supported protocols (OpenAI-style / Anthropic-style), model lists with context/output-token capabilities, and (where available) a usage/quota lookup — for building AI gateways, routers, or any tool that needs to know "what providers and models exist and how do I talk to them."

Originally extracted from ModelPlane.dev's AI gateway so the catalog can grow independently: adding a new provider or fixing a stale model list shouldn't require touching gateway internals, and anyone is welcome to contribute.

Install

npm install llm-provider-catalog

Usage

import { listCatalogProviders, getCatalogProvider } from 'llm-provider-catalog';

const providers = listCatalogProviders();

const zai = getCatalogProvider('zai');
console.log(zai?.baseURLs.openai); // https://api.z.ai/api/paas/v4

const zaiCoding = getCatalogProvider('zai-coding');
if (zaiCoding?.fetchUsage) {
  const usage = await zaiCoding.fetchUsage(apiKey);
  console.log(usage.quotas);
}

What's in a CatalogProvider

export interface CatalogProvider {
  id: string; // unique catalog id, e.g. "openai" or "zhipu"
  label: string; // human-readable display name
  brandId?: string; // groups multiple catalog entries under one brand (e.g. minimax coding plan + standard plan)
  codingPlan?: boolean; // flags a coding-plan-specific entry (separate quota/billing from the standard plan)
  provider: string; // underlying protocol-transform identifier a gateway would route on
  protocols: ApiProtocol[]; // 'openai' | 'anthropic' — which request/response shapes this provider accepts
  baseURLs: Partial<Record<ApiProtocol, string>>; // base URL per supported protocol
  models: CatalogModel[]; // model id, display label, and capability (context window, max output tokens, tokenizer, modalities, features)
  embeddingModels?: CatalogEmbeddingModel[]; // same shape, for providers with an embeddings endpoint
  fetchUsage?(apiKey: string): Promise<ProviderUsageResult>; // optional balance/quota lookup
  fetchModels?(apiKey?: string): Promise<FetchModelsResult>; // optional live model-id lookup
  fetchEmbeddingModels?(apiKey?: string): Promise<FetchEmbeddingModelsResult>; // optional live embedding-model lookup
}

fetchModels/fetchEmbeddingModels are opt-in, per-provider live lookups against the provider's own model-list API — fetchModels returns just the model ids currently available to a key (most providers' /models endpoints don't expose capability data), while fetchEmbeddingModels returns full CatalogEmbeddingModel objects where the provider's endpoint publishes capability info (currently only OpenRouter).

ModelCapability also carries optional multimodal info: modalities: { input: Modality[]; output: Modality[] } (Modality is 'text' | 'image' | 'audio' | 'video' | 'file'), omitted entirely for plain text-in/text-out models, and features: { toolUse?, structuredOutputs?, toolUseWithVision?, codeExecution?, webSearch? } for well-known cross-vendor capability flags. Both are populated from live data for OpenRouter/Novita AI (their /models APIs expose this directly); for other providers they're filled in only when the vendor's own docs explicitly confirm it — an unset field means "unverified," not "unsupported."

Full field-level type definitions live in src/types.ts.

Live testing

npm run test:live (see CONTRIBUTING.md) exercises each catalog provider against its real API using credentials in tests/live/.usage-creds.json (gitignored, copy from .usage-creds.json.example). Cases for a provider are skipped automatically when its credential is missing.

As of this writing, these providers have not been exercised against a live account (no credential available) and should be treated as unverified until someone runs them with real keys:

• anthropic
• google (Gemini)
• minimax-global (Minimax standard plan)
• minimax-china (Minimax standard plan, China)
• zhipu (Zhipu / BigModel, China)
• zhipu-coding (Zhipu Coding Plan, China)
• minimax-coding-china (Minimax Coding Plan, China)
• kimi (Kimi / Moonshot AI, International)
• kimi-china (Kimi / Moonshot AI, China)
• kimi-coding (Kimi Code)

Contributing

Adding a new provider, fixing a stale model list, or fixing a bug — see CONTRIBUTING.md. No API keys or live network calls are required to add a provider; only implementing/changing fetchUsage benefits from live credentials, and even that ships with a mocked unit test path.

License

MIT