Zero-Key GCP Setup: Binding a Service Account to Compute Engine for Vertex AI

June 2, 2026 • 6 min read

If you're running Indexu on Google Compute Engine, you can skip JSON key files entirely. By binding a service account directly to your VM, authentication happens automatically through the metadata server — no keys to store, rotate, or leak.

Why Service Binding Is Better Than JSON Keys

When you attach a service account to a GCE instance, the VM inherits that identity. Any process running on the VM can automatically authenticate using Google's Application Default Credentials (ADC) — no key file needed.

	JSON Key File	Service Binding (GCE)
Secret to manage	JSON key file on disk or in env vars	None
Rotation	Manual key rotation required	Automatically handled by GCP
Leak risk	Key can be used from anywhere	Tied to the VM — useless elsewhere
Access revocation	Delete the key in IAM	Detach from VM or stop the instance
Works outside GCP?	Yes	No — VM must run on GCE

Step 1: Create the Scoped Service Account

The service account setup is identical to the JSON key approach — same least-privilege roles, same scope. The difference is how the VM uses it.

1. Go to IAM & Admin → Service Accounts
2. Click Create Service Account
3. Name it indexu-vertexai-gce
4. Assign exactly these roles:

Role	What It Grants
roles/aiplatform.user	Access Vertex AI models for prediction
roles/storage.objectUser	Read/write objects in Cloud Storage buckets

Click Done. Do not create any keys — you won't need them.

Step 2: Bind the Service Account to Your VM

Service account binding happens at the VM instance level. You have two options:

Option A: Set During VM Creation

When creating a new Compute Engine instance via the console:

1. Expand the Identity and API access section
2. Under Service account, select your indexu-vertexai-gce account
3. Under Access scopes, choose Set access for each API
4. Set Cloud Storage to Read Write
5. Click Create

Option B: Update an Existing VM

If the VM is already running, you can update it via the gcloud CLI:

gcloud compute instances set-service-account INSTANCE_NAME \
  --service-account=indexu-vertexai-gce@PROJECT_ID.iam.gserviceaccount.com \
  --scopes=cloud-platform,storage-rw \
  --zone=us-central1-a

The cloud-platform scope grants broad API access, but the service account's IAM roles still enforce the actual permissions. The effective access is the intersection of the VM's access scopes and the service account's IAM roles.

Step 3: How Authentication Works (No Config Needed)

Once the service account is bound, authentication is fully automatic. Here's what happens under the hood:

1. Your application code calls a Google Cloud API (Vertex AI or Storage)
2. The Google client library detects it's running on GCE
3. It queries the metadata server at http://metadata.google.internal
4. The metadata server returns a short-lived OAuth 2.0 access token for the bound service account
5. The client library attaches this token to the API request

You can verify this works by SSHing into your VM and running:

# Fetch an access token from the metadata server
curl "http://metadata.google.internal/computeMetadata/v1/instance/service-accounts/default/token" \
  -H "Metadata-Flavor: Google"

This returns a JSON response with an access_token field — the same token the client libraries use automatically.

Step 4: Set Up Vertex AI Quotas

Quota setup is the same regardless of how you authenticate. If you haven't already requested quota, do it now:

1. Go to IAM & Admin → Quotas and search for Vertex AI
2. Request increases for these metrics:

Quota	Recommended
LLM online prediction requests per minute	100
Base model online prediction tokens per minute	300,000
Online prediction requests per minute	60
Concurrent online prediction requests	5

Select each quota row, click Edit Quota, enter your desired limit, and provide a brief justification. Approval typically takes a few hours.

Step 5: Firewall and Network Considerations

Since your VM is the only entity that can use this service account, network security is simpler. However, there are a few things to check:

Egress Firewall Rules

Your GCE instance needs outbound access to these Google API endpoints:

• aiplatform.googleapis.com (Vertex AI)
• storage.googleapis.com (Cloud Storage)
• oauth2.googleapis.com (token endpoint)
• metadata.google.internal (metadata server — always accessible internally)

If you use restrictive egress rules, make sure these endpoints are allowed. The easiest path is to allow outbound HTTPS (tcp:443) to 0.0.0.0/0 for your VM's subnet — Google APIs are internet-facing.

Private Google Access (Optional but Recommended)

For an extra layer of security, enable Private Google Access on your VPC subnet. This routes API traffic through Google's backbone instead of the public internet:

1. Go to VPC network → VPC networks
2. Click your VPC, then the subnet your VM is in
3. Set Private Google Access to On

With this enabled, your VM never needs a public IP or internet gateway for Google API traffic.

Step 6: Verify the Setup

SSH into your GCE instance and run these tests to confirm everything is wired correctly:

# 1. Confirm the service account identity
gcloud auth list
# Expected: indexu-vertexai-gce@PROJECT_ID.iam.gserviceaccount.com (active)

# 2. Test Vertex AI access via the metadata token
ACCESS_TOKEN=$(curl -s "http://metadata.google.internal/computeMetadata/v1/instance/service-accounts/default/token" \
  -H "Metadata-Flavor: Google" | python3 -c "import sys,json; print(json.load(sys.stdin)['access_token'])")

curl -s -X POST \
  -H "Authorization: Bearer $ACCESS_TOKEN" \
  -H "Content-Type: application/json" \
  "https://us-central1-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/us-central1/publishers/google/models/gemini-2.5-flash:streamGenerateContent" \
  -d '{"contents":[{"role":"user","parts":[{"text":"Say hello"}]}]}'

# 3. Test Cloud Storage access
gsutil ls gs://your-bucket-name/

# 4. Verify no key file is needed
echo $GOOGLE_APPLICATION_CREDENTIALS
# Expected: (empty)

If all three succeed and GOOGLE_APPLICATION_CREDENTIALS is empty, your setup is complete and fully keyless.

Summary

With service account binding, you've achieved a setup that is:

• Keyless — no JSON file to store, rotate, or protect
• Scoped — only aiplatform.user and storage.objectUser permissions
• VM-bound — credentials are inherently tied to the instance and can't be used elsewhere
• Zero-config — Indexu picks up the service account automatically via ADC

This is the recommended setup for any Indexu deployment running on Google Compute Engine. Compared to the JSON key approach, it eliminates the most common security concern — leaked credentials — while requiring less ongoing maintenance.

Need help deploying Indexu on GCE?
Contact us and we'll help you configure the instance, service account, and networking for a production-ready setup.