LLM Provider APIs

Pass one or more provider API keys to Nua.direct() and Nuabase calls the provider’s API directly over HTTPS, from your process. There is no CLI to install and nothing in between you and the provider.

Configuration

import { Nua, z } from 'nuabase';

const nua = Nua.direct({
  providers: {
    groq: { apiKey: process.env.GROQ_API_KEY! },
    cerebras: { apiKey: process.env.CEREBRAS_API_KEY! },
    gemini: { apiKey: process.env.GEMINI_API_KEY! },
    openrouter: { apiKey: process.env.OPENROUTER_API_KEY! },
    deepseek: { apiKey: process.env.DEEPSEEK_API_KEY! },
  },
  model: { alias: 'fast' }, // the default
});

Configure only the providers you have keys for. A single provider is enough.

Choosing a model

Set a default model on the client, and override it per call. A model is either an alias or an explicit provider and model:

// By alias: uses the first provider in the alias's list that you have a key for
await nua.get('Extract the company name', {
  input: text,
  output: { name: 'companyName', schema: z.string() },
  model: { alias: 'sonnet' },
});

// By provider and model
await nua.get('Extract the company name', {
  input: text,
  output: { name: 'companyName', schema: z.string() },
  model: { provider: 'openrouter', model: 'z-ai/glm-5.3-flash' },
});

With no model set anywhere, Nuabase uses { alias: 'fast' }, which tries Cerebras, then Groq, then OpenRouter. If you have a key for none of an alias’s providers, or name a provider you haven’t configured, the call returns success: false.

See Models & Providers for every alias and provider.

Runtimes

The main nuabase entry point imports no Node.js built-ins, so a client configured with providers runs in Node.js and in edge runtimes such as Cloudflare Workers and Vercel Edge.

Your API keys go in the request to the provider, so run these calls on a server or in an edge function, never in code shipped to a browser.

Results

if (result.success) {
  result.model;         // "openrouter:z-ai/glm-5.3-flash"
  result.meta.engine;   // "http"
  result.meta.provider; // "openrouter"
  result.meta.model;    // "z-ai/glm-5.3-flash"
  result.usage;         // { promptTokens, completionTokens, totalTokens, promptCacheReadTokens?, promptCacheWriteTokens? }
}

promptCacheReadTokens and promptCacheWriteTokens are present when the provider reports them. See prompt caching for how to get cache hits.

Requests and retries

Each call makes one HTTPS request to the selected provider, plus up to 2 retries on transient errors or output that fails validation. A 401, 403, or 404 response (a bad API key, no access, or an unknown model) fails at once. Rate-limit and quota responses (429) are retried like other errors, and Nuabase sets no request timeout of its own.

Moving between engines

A client built with providers and one built with localAgent have the same get() and list() methods and return the same result shape. A common pattern is to develop against your coding agent and deploy with an API key:

import { Nua } from 'nuabase';

const nua = process.env.OPENROUTER_API_KEY
  ? Nua.direct({ providers: { openrouter: { apiKey: process.env.OPENROUTER_API_KEY } } })
  : Nua.direct({ localAgent: await (await import('nuabase/local-agent')).findLocalAgent() });

The dynamic import keeps nuabase/local-agent, which needs Node.js, out of code paths that run in edge runtimes.

Model aliases are the one difference to watch: a per-call model: { alias: 'sonnet' } works on both an OpenRouter client and Claude Code, but { alias: 'fast' } has no local agent equivalent, and provider models can’t run on a local agent.