Google Vertex AI

Anthropic models served from Google Cloud Vertex AI.

Quick start

agents harness create
# pick vertex, fill prompts, run smoke test
agents run my-profile "hello"

Required values

Var	Where to get it
GCP project id	gcloud config get-value project
Region	A Vertex region where the model is available (e.g. us-east5, europe-west1)
Service-account key	GCP IAM → Service Accounts → Keys → JSON; path goes in GOOGLE_APPLICATION_CREDENTIALS
Model id	Vertex model garden, e.g. claude-sonnet-4-5@20250929

Generated profile shape

name: my-profile
host: { agent: claude }
env:
  # static vars (always set):
  CLAUDE_CODE_USE_VERTEX: "1"
  # vars wizard collects:
  ANTHROPIC_VERTEX_PROJECT_ID: my-gcp-project
  CLOUD_ML_REGION: us-east5
  GOOGLE_APPLICATION_CREDENTIALS: /Users/me/.config/gcloud/vertex-sa.json
  ANTHROPIC_MODEL: claude-sonnet-4-5@20250929
  ANTHROPIC_SMALL_FAST_MODEL: claude-haiku-4-5@20251001
# no auth: block — Vertex uses GOOGLE_APPLICATION_CREDENTIALS, not a bearer token

Known caveats

Auth is a service-account JSON file, not a bearer token. GOOGLE_APPLICATION_CREDENTIALS must be the path to a JSON key file on disk. There is no ANTHROPIC_AUTH_TOKEN for Vertex — the Google auth library mints short-lived access tokens from the service-account key on every request. As a result the profile's auth: block is typically omitted.

Region-specific model availability. claude-sonnet-4-6 and claude-opus-4-7 are not in every Vertex region. Check the current availability matrix in the Vertex Model Garden before pinning CLOUD_ML_REGION — us-east5 and europe-west1 are the broadest, but newer models often land in us-east5 first.

Model id includes a date suffix. Vertex uses claude-sonnet-4-5@20250929, not the bare claude-sonnet-4-5. The @<date> pins a specific snapshot.

Troubleshooting

Symptom	Cause	Fix
Could not load the default credentials	GOOGLE_APPLICATION_CREDENTIALS unset or pointing at a missing file	Re-export the path; verify with cat $GOOGLE_APPLICATION_CREDENTIALS | jq .type returning "service_account"
Publisher Model ... not found	Model not available in this region	Switch CLOUD_ML_REGION, or pick a model id present in your region
PERMISSION_DENIED: aiplatform.endpoints.predict	Service account missing the Vertex AI User role	Grant roles/aiplatform.user on the project to the service account
429 quota errors	Per-project / per-region Vertex quota exhausted	Request a quota increase, or shift region