Setup

Install the Gemini adapter next to the core runtime:

pnpm add @anvia/core @anvia/gemini

Configure the Gemini API

Set the API key in the server environment:

GEMINI_API_KEY="your-api-key"

Create the client in a server-only module:

import { GeminiClient } from '@anvia/gemini'

export const gemini = new GeminiClient({
  apiKey: process.env.GEMINI_API_KEY,
})

export const completionModel = gemini.completionModel(
  'gemini-2.5-flash',
)

Construction fails when the API key is missing or empty. Validate environment configuration at startup so the deployment fails before it accepts traffic.

Use the model with an agent

import { AgentBuilder } from '@anvia/core'
import { completionModel } from './gemini'

export const supportAgent = new AgentBuilder(
  'support',
  completionModel,
)
  .instructions(
    'Answer support questions clearly. Use tools for account data.',
  )
  .defaultMaxTurns(4)
  .build()

The agent remains provider-neutral. Provider selection stays in the model module, making it easier to test the workflow with a fake model or replace the deployment later.

Inject an existing Google client

Pass an existing GoogleGenAI instance when the application needs to own its construction:

import { GoogleGenAI } from '@google/genai'
import { GeminiClient } from '@anvia/gemini'

const google = new GoogleGenAI({
  apiKey: process.env.GEMINI_API_KEY,
})

export const gemini = new GeminiClient({ client: google })

Install @google/genai directly when application code imports it. Authentication and SDK lifecycle are then the application's responsibility.

Keep credentials out of the browser

Browser code should call an application route that owns the agent or completion request. Never expose a Gemini key through public environment variables, serialized page data, or browser-side SDK construction.

Continue to Vertex AI when the deployment uses Google Cloud IAM instead of an API key.