Google Vertex AI

Vertex AI serves the same Gemini models over the same wire dialect as the Gemini API — the difference is governance. Inference runs inside your Google Cloud project, billed to your account, subject to your IAM, quotas, and data-residency controls. If your organization requires Google Cloud to be the front door, this is the provider; if not, the plain Gemini API is one env var instead of three.

Setup

Vertex authenticates with a Google Cloud access token and needs a project:

export VERTEX_ACCESS_TOKEN=$(gcloud auth print-access-token)
export VERTEX_PROJECT_ID="my-gcp-project"      # or GOOGLE_CLOUD_PROJECT
export VERTEX_LOCATION="global"                # optional — this is the default

GOOGLE_CLOUD_PROJECT works in place of VERTEX_PROJECT_ID; VERTEX_LOCATION defaults to global and only needs setting if you pin a regional endpoint.

The shell function, concretely — mint a fresh token per invocation and forget about expiry:

# ~/.zshrc or ~/.bashrc
stella-vertex() {
  VERTEX_ACCESS_TOKEN="$(gcloud auth print-access-token)" stella "$@"
}

stella-vertex --model vertex/gemini-3-pro run "audit the IAM bindings module"

Each call gets a token good for about an hour — enough for the run it starts, and the next call mints its own.

Selecting models

The provider id is vertex; the default model is gemini-3-pro, and the lineup mirrors the Gemini API.

stella --model vertex/gemini-3-pro run "audit the IAM bindings module"

Pinning with --model vertex/… matters here, because of detection order.

Last in auto-detection — on purpose

Vertex sits last in auto-detection (alongside Bedrock), deliberately: VERTEX_ACCESS_TOKEN and ambient Google Cloud credentials are routinely present in shells for reasons that have nothing to do with LLM inference, and a generic cloud credential must never hijack provider selection. If you want Vertex, say so — --model vertex/gemini-3-pro, or set it as a default in settings.json:

{
  "agent_engine_config": {
    "default_model": "vertex/gemini-3-pro"
  }
}

The pin sits between the --model flag and auto-detection — an explicit flag still wins, but Vertex no longer loses to whichever direct-vendor key is also in the shell.

Thinking

Identical to the Gemini API: Stella's effort setting maps to thinkingLevel (low/high), and Gemini 3 cannot fully disable thinking — reasoning: "off" yields the minimum level, not zero.

In the pipeline

Same models, same roles as Gemini — a 1M-window worker or judge — chosen when inference must stay inside your cloud perimeter. Route it per agent with agent_engine_config; see Examples for profile patterns.

Catalog pricing

Catalog list prices from Stella's seed catalog mirror the Gemini API rates — your actual Vertex billing is set by your Google Cloud agreement; verify there. Your local catalog re-syncs from models.dev via stella models refresh.