@mayurajs/provider-google

Google Gemini models for Mayura, through the official @google/genai SDK and the Gemini API.

npm install mayura @mayurajs/provider-google

import { createModels, defineAgent, z } from 'mayura';
import { google } from '@mayurajs/provider-google';

const models = createModels({
  providers: [google({ apiKey: process.env.GEMINI_API_KEY! })],
  prices: { 'google/gemini-3.5-flash': { inputMicrosPerMillionTokens: 1_500_000, outputMicrosPerMillionTokens: 9_000_000 } },
  maxCallCostMicros: 50_000,
});

const agent = defineAgent({
  id: 'helper', version: '1', instructions: 'Answer in one sentence.', model: models.model('google/gemini-3.5-flash'), tools: [],
  input: z.object({ question: z.string() }), output: z.object({ answer: z.string() }),
});
// Grant it as model:google/gemini-3.5-flash.

• Credentials are options: nothing is read from the environment (not GEMINI_API_KEY, GOOGLE_API_KEY, GOOGLE_GEMINI_BASE_URL or the Vertex AI settings).
• The SDK's own retries are off; the registry's retry option retries and charges every attempt.
• Tool parameters and structured output use strict JSON Schema; streaming reports output text as it arrives; the model's thoughts and thought signatures are kept for the next call and never released.
• Thinking tokens are charged as output, and cached input at the full input rate. Media is sent as bytes.
• A run's conversation history (capabilities.history) goes first, in Gemini's turn order: turns of one role in a row share a turn, and a history that starts with an answer gets a user turn holding null before it. History has no thought signatures; Gemini checks them only in the current turn.
• catalog holds the Gemini API's list prices on its date, used only with prices: 'catalog'. Gemini 3.8 Flash is listed at the price announced from January 1, 2027, so a budget never undercounts.

See the model registry guide. Apache-2.0.

RAG embeddings

This provider exports googleEmbeddings for scoped RAG ingestion and retrieval. See the RAG guide for explicit permissions, cost ceilings and sensitivity profiles. Development cases are authored; validation is deferred.