@amanm/vertex-maas

Vertex MaaS provider for the Vercel AI SDK.

This provider targets Vertex AI models exposed via OpenAI-compatible endpoints and authenticates using Google Application Default Credentials (ADC) (service account JSON via GOOGLE_APPLICATION_CREDENTIALS, workload identity, etc.).

Requirements

• Node.js runtime (not Edge)
• ADC configured (commonly via GOOGLE_APPLICATION_CREDENTIALS)
• Set GOOGLE_VERTEX_PROJECT and GOOGLE_VERTEX_LOCATION (or pass project/location to createVertexMaaS)

Usage

import { generateText } from 'ai';
import { createVertexMaaS } from '@amanm/vertex-maas';

const vertex = createVertexMaaS({
  project: process.env.GOOGLE_VERTEX_PROJECT!,
  location: process.env.GOOGLE_VERTEX_LOCATION!, // e.g. "global" or "us-central1"
});

const result = await generateText({
  model: vertex('glm-4.7'),
  prompt: 'Explain Amdahl\\'s Law in simple terms.',
});

Settings

• baseURL: Optional. If set, used directly.
• project / location: Used to compute the Vertex OpenAI-compatible base URL if baseURL is not provided.
• apiKey: Optional. If set, uses a static Authorization: Bearer <apiKey> header (no ADC).
• googleAuthOptions: Optional. Passed to google-auth-library GoogleAuth.
• opencodeCompat: Optional. If true, maps finish.finishReason from { unified, raw } to unified in doStream() for OpenCode.

OpenCode

OpenCode currently expects the model stream finishReason to be a string.

Set opencodeCompat: true in your OpenCode provider options:

{
  "provider": {
    "vertexmaas": {
      "npm": "@amanm/vertex-maas",
      "options": {
        "project": "common-project-by-aman",
        "location": "global",
        "opencodeCompat": true
      },
      "models": {
        "glm-4.7-maas": { "id": "zai-org/glm-4.7-maas" }
      }
    }
  }
}