vertex

Vertex AI — Gemini on Google Cloud with project + region routing and OAuth2 tokens.

import { vertex } from '@agentskit/adapters'

const adapter = vertex({
  project: process.env.GCP_PROJECT!,
  region: 'us-central1',
  model: 'gemini-2.5-pro',
  accessToken: process.env.GCP_ACCESS_TOKEN!,
})

accessToken can be a string or a () => Promise<string> minted on demand. The adapter intentionally does not depend on google-auth-library; mint tokens with gcloud auth print-access-token, the auth library, or your own Workload Identity flow.

#Options

Option	Type	Default
`project`	`string`	required
`region`	`string`	required (e.g. `us-central1`)
`model`	`string`	required (e.g. `gemini-2.5-pro`)
`accessToken`	`string \| (() => string \| Promise<string>)`	required
`publisher`	`string`	`google` (set for partner publishers like `anthropic`)
`retry`	`RetryOptions`	inherited

#Capabilities

{ streaming: true, tools: true, multiModal: true, usage: true, reasoning: model.includes('pro') }. Request shape mirrors the public Gemini API; auth is OAuth2 instead of an API key.

#Why vertex

GCP-native deployments that need Gemini behind project IAM.
Partner publishers (Anthropic on Vertex, Mistral on Vertex) under one auth model.
Workload Identity / Service Account auth.

#Caveats

This adapter does not mint tokens. Bring accessToken from your auth flow (function form is recommended in long-lived processes — tokens expire after ~1 hour).
Knowledge Bases / Grounding go in @agentskit/rag (not in v1).

#Env

Var	Purpose
`GCP_PROJECT`	GCP project id
`GCP_ACCESS_TOKEN`	OAuth2 token (or mint dynamically in `accessToken`)

Providers overview · Choosing an adapter · gemini · bedrock

Explore nearby

✎ Edit this page on GitHub·Found a problem? Open an issue →·How to contribute →