Vertex AI

Vertex AI hosts Gemini, Anthropic Claude, Meta Llama, and Mistral models on your own GCP project. Use it when you want Google-managed inference billed against your GCP account, regional control, or models that aren't on AI Studio.

Type	Models
`vertex-google`	Gemini models
`vertex-anthropic`	Claude models
`vertex-meta`	Llama models
`vertex-mistralai`	Mistral models

All four use the same auth and URL pattern. Pick the one that matches the model family you want.

Prerequisites

A GCP project with billing enabled.

The Vertex AI API enabled on that project:

gcloud services enable aiplatform.googleapis.com --project YOUR_PROJECT_ID

A region — e.g. us-central1, europe-west4. Each model is only available in some regions; check the Vertex AI model availability matrix.

The backend URL always follows this shape:

https://{REGION}-aiplatform.googleapis.com/v1/projects/{PROJECT_ID}/locations/{REGION}

Auth method 1 — Service account JSON (recommended for servers)

Create a service account with the roles/aiplatform.user role, download a JSON key, and load it through an env var:

export VERTEX_SA_JSON=$(cat /path/to/service-account.json)

contenox backend add vertex --type vertex-google \
  --url "https://us-central1-aiplatform.googleapis.com/v1/projects/$GOOGLE_CLOUD_PROJECT/locations/us-central1" \
  --api-key-env VERTEX_SA_JSON

contenox config set default-model gemini-2.5-flash
contenox config set default-provider vertex-google

Contenox reads the JSON from the named env var at request time, so the key never lands in the config file on disk.

Auth method 2 — Application Default Credentials (CLI / dev only)

ADC reuses your gcloud login. It's the fastest way to try Vertex but expires when the refresh token is revoked (see renewal below).

gcloud config set project YOUR_PROJECT_ID
gcloud services enable aiplatform.googleapis.com
gcloud auth application-default login
gcloud auth application-default set-quota-project YOUR_PROJECT_ID

contenox backend add vertex --type vertex-google \
  --url "https://us-central1-aiplatform.googleapis.com/v1/projects/YOUR_PROJECT_ID/locations/us-central1"

contenox config set default-model gemini-2.5-flash
contenox config set default-provider vertex-google

Omit --api-key-env and Contenox falls back to ADC.

Important

set-quota-project is required. Without it, every Vertex AI call returns 403 SERVICE_DISABLED — even if you already ran gcloud config set project.

Renewing credentials

Vertex tokens are short-lived (~1 hour). Contenox refreshes them automatically on every request, so day-to-day you never deal with the access token itself — what you renew is the credential the token is minted from.

Symptom: `vertex AI token refresh: oauth2: "invalid_grant"`

This means ADC tried to mint a fresh access token from its refresh token and Google rejected the refresh token. Causes, most common first:

You haven't run gcloud auth application-default login in a long time and the refresh token aged out (Google rotates them).
You changed your Google account password.
The account's session was revoked (admin action, security policy, or you ran gcloud auth application-default revoke).
The ADC file at ~/.config/gcloud/application_default_credentials.json was deleted or replaced by a different account.

Fix:

gcloud auth application-default login
gcloud auth application-default set-quota-project YOUR_PROJECT_ID

No contenox restart needed — the next request picks up the new credentials. If you removed the ADC file by hand, also re-run set-quota-project.

Symptom: `vertex AI service account token: ...`

The service account JSON in VERTEX_SA_JSON is invalid, the key was disabled in GCP, or the service account was deleted. Generate a new key:

gcloud iam service-accounts keys create new-key.json \
  --iam-account=vertex-runner@YOUR_PROJECT_ID.iam.gserviceaccount.com

export VERTEX_SA_JSON=$(cat new-key.json)

Then restart the shell / process that holds the env var so Contenox sees the new value.

Rotating a service account key on a schedule

Service account keys don't expire on their own — rotate them yourself. Typical pattern:

# Create a new key
gcloud iam service-accounts keys create new-key.json \
  --iam-account=vertex-runner@YOUR_PROJECT_ID.iam.gserviceaccount.com

# Swap the env var (in your secret store / systemd unit / k8s secret)
export VERTEX_SA_JSON=$(cat new-key.json)

# Disable the old key once requests are flowing on the new one
gcloud iam service-accounts keys list \
  --iam-account=vertex-runner@YOUR_PROJECT_ID.iam.gserviceaccount.com

gcloud iam service-accounts keys delete OLD_KEY_ID \
  --iam-account=vertex-runner@YOUR_PROJECT_ID.iam.gserviceaccount.com

Troubleshooting

Error	Cause	Fix
`oauth2: "invalid_grant"`	ADC refresh token revoked / aged out	`gcloud auth application-default login`
`403 SERVICE_DISABLED`	Quota project not set, or Vertex AI API not enabled	`gcloud services enable aiplatform.googleapis.com` and `gcloud auth application-default set-quota-project ...`
`403 PERMISSION_DENIED` on a service account	Missing `roles/aiplatform.user`	Grant the role on the project
`404` on the model	Model not available in your region	Pick a region from the model availability matrix and recreate the backend with the right URL
`unreachable: vertex-google list models: ...` in `contenox model list`	Same as `invalid_grant` above — the catalog fetch refreshes tokens too	Renew per the section above; the backend stays registered