Multi-tenant AI assistant platform on Amazon EKS. Each user gets an isolated, private AI workspace powered by Amazon Bedrock -- zero API keys, zero shared data.
Deploy in ~30 minutes. Scale to 500 users. Pay only for what you use.
Experimental — This project is provided for experimentation and learning purposes only. It is not intended for production use. APIs, architecture, and configuration may change without notice. See Security for details.
- One tenant per user -- isolated namespace, PVC, network policy, IAM role
- Zero API keys -- LLM access via Amazon Bedrock + Pod Identity
- Scale to zero -- KEDA scales idle pods to 0; cold start in 15-30s
- 3-layer origin protection -- internet-facing ALB with CF-only SG + AWS WAF + HTTPS
- Custom auth UI -- branded login/signup on your domain (no Amazon Cognito Hosted UI)
- Self-service signup -- Amazon Cognito + AWS Lambda auto-provisions tenants
- ArgoCD ApplicationSet managed -- ApplicationSet generates per-tenant ArgoCD Applications; each syncs Helm chart with tenant-specific values (details)
- Cost control -- per-tenant monthly budget with per-model pricing alerts
- Graviton ARM64 -- 20% cheaper compute with t4g instances
Path-based routing via Gateway API: claw.example.com/t/<tenant>/ -- one domain, one ALB, no wildcard DNS needed.
Internet
|
+- claw.your-domain.com --> Amazon CloudFront (single distribution)
| / -> S3 (auth UI) [CDK-managed]
| /t/* -> Internet-facing ALB [post-deploy.sh]
| (CF-only SG + AWS WAF) -> Amazon EKS Pod
+- Outbound only: Amazon EKS Pod --> NAT Gateway (HA) --> Internet
Amazon EKS Cluster
| Managed Node Group (Graviton ARM64) + Karpenter (arm64 spot)
| Add-ons: ALB Controller, EBS CSI, Amazon EFS CSI, Pod Identity, CloudWatch Insights
| KEDA HTTP Add-on
|
+-- namespace: openclaw-{tenant}
| ApplicationSet-managed (ArgoCD):
| Namespace PVC (Amazon EFS)
| ArgoCD Application ServiceAccount (Pod Identity)
| ReferenceGrant (keda ns) Deployment + Service + ConfigMap
| HTTPRoute + TGC + NetworkPolicy + PDB + KEDA HSO
| ArgoCD Application (in argocd namespace, points to helm/charts/openclaw-platform)
git clone https://github.com/snese/sample-openclaw-multi-tenant-platform.git
cd sample-openclaw-multi-tenant-platform
./setup.shsetup.sh checks prerequisites, prompts for configuration, and deploys everything (~20 min).
- AWS CLI v2 + configured profile
- AWS CDK v2 >= 2.1114.1 (
npm install -g aws-cdk) - Docker or Finch (running — required for AWS CDK asset bundling). If using Finch, set
export CDK_DOCKER=finchbefore running CDK commands - kubectl + Helm 3, Node.js 22+
- Python 3 — required for CloudFront configuration management
Important: Set
allowedEmailDomainsincdk.jsonto restrict who can sign up (e.g.,your-company.com). Without this, anyone with a valid email can create a tenant. To disable self-signup entirely, setselfSignupEnabledtofalseincdk.json.
- (Optional) Route53 hosted zone + ACM certificate in us-east-1. Without a custom domain, the platform uses the CloudFront default domain (
xxxxxx.cloudfront.net)
cp cdk/cdk.json.example cdk/cdk.json
# Edit cdk/cdk.json -- fill in context values (see cdk.json.example for full list)cd cdk && npm install
cdk bootstrap # Only needed once per account/region
cd ..
REGION=us-west-2 bash scripts/deploy-all.shdeploy-all.sh runs all steps automatically: CDK deploy (~20 min), ArgoCD, platform resources, KEDA, CloudFront ALB origin + Route53, and Cognito verification.
Manual steps: If you prefer to run each step individually, see
scripts/deploy-all.shfor the sequence.
ECR Pull-Through Cache (optional): For production, you can enable ECR pull-through cache to avoid GHCR rate limits. Set
ghcrCredentialArnincdk.json-- seecdk.json.examplefor details.
Sign up at https://<your-domain>/auth/ with an email matching your allowedEmailDomains.
Note: Auth UI is deployed automatically by AWS CDK (
BucketDeployment). If you need to manually re-deploy or override config, run./scripts/deploy-auth-ui.sh.
./scripts/create-tenant.sh <name> [options] # Create
./scripts/delete-tenant.sh <name> # Delete (with confirmation)
./scripts/verify-tenant.sh <name> # Health check
./scripts/check-all-tenants.sh # Check all tenants
./scripts/backup-tenant.sh <name> <bucket> # Backup to Amazon S3
./scripts/restore-tenant.sh <name> <s3-path> # Restore from Amazon S3
./scripts/admin-list-tenants.sh # List tenants + cost| Layer | Control |
|---|---|
| Edge | Amazon CloudFront + AWS WAF (CLOUDFRONT scope, edge protection) |
| Signup | AWS WAF Bot Control (opt-in) + email domain restriction + rate limiting |
| Network | Internet-facing ALB with CF-only SG (pl-82a045eb) + AWS WAF + HTTPS |
| Auth | Amazon Cognito signup + local token auth + 3-layer origin protection |
| Tenant | Namespace isolation + NetworkPolicy + ABAC |
| Secrets | exec SecretRef -- fetched on-demand, never persisted |
| LLM | Amazon Bedrock via Pod Identity -- zero API keys |
| Cost | Per-tenant monthly budget with per-model pricing |
| Data | PVC persists across scale-to-zero (Amazon EFS, multi-AZ) |
| Audit | CloudTrail + Amazon S3 + Athena + Amazon EKS control plane logging |
| Resource | 3 tenants | 100 tenants |
|---|---|---|
| Amazon EKS control plane | ~$73 | ~$73 |
| EC2 (3x t4g.large system + Karpenter spot) | ~$93 | ~$93-180 |
System nodegroup uses 3x t4g.large to run platform components (ArgoCD, KEDA, ALB Controller, Karpenter, CloudWatch, GuardDuty). These components request ~8.5 vCPU total (upstream Helm/addon defaults), though actual usage is ~15-25%. Karpenter provisions additional spot nodes for tenant pods on demand. | Amazon EFS (per actual usage) | ~$0.15 | ~$75 | | ALB + NAT (x2) + Amazon CloudFront + AWS WAF | ~$60 | ~$65 | | CloudWatch + AWS Lambda + Amazon S3 |
$15 |$331-418/mo** |$20 | | Amazon Bedrock | varies | varies | | Total (infra) | **$243/mo** | **
KEDA scale-to-zero active. EC2 scales with concurrent usage, not total tenants.
| Doc | Description |
|---|---|
| Architecture | ApplicationSet + ArgoCD flow, tenant lifecycle |
| Security | 10-layer security model |
| Amazon EKS Cluster | Cluster, nodegroups, Karpenter, add-ons |
| Networking | VPC, Amazon CloudFront, ALB, AWS WAF |
| Auth | Amazon Cognito, custom UI, AWS Lambda triggers |
| Scaling | KEDA scale-to-zero, cold start |
| GitOps | ArgoCD + ApplicationSet, tenant lifecycle |
| Admin Guide | Deploy, manage, monitor |
| User Guide | Signup, login, daily use |
| Troubleshooting | Common issues and fixes |
+-- auth-ui/ # Custom login/signup (Amazon S3 + Amazon CloudFront)
+-- cdk/ # AWS CDK infrastructure (TypeScript)
| +-- lib/eks-cluster-stack.ts
| +-- lambda/ # Pre-signup, Post-confirmation, Cost-enforcer
| +-- cdk.json.example # Config template
+-- helm/ # Helm chart (source of truth, synced by ArgoCD)
| +-- charts/openclaw-platform/ # Tenant K8s resources (Deployment, Service, etc.)
| +-- gateway.yaml # Gateway API resource
| +-- tenants/values-template.yaml # Example values
+-- helm/applicationset.yaml # ArgoCD ApplicationSet (multi-tenant generator)
+-- docs/ # Architecture, security, components, operations
+-- scripts/ # 20+ operations scripts
+-- .github/workflows/ci.yml # CI pipeline
REGION=us-east-1 bash scripts/destroy-all.shdestroy-all.sh removes everything in reverse order: tenants → KEDA → Gateway (ALB) → ArgoCD → CDK destroy → log groups. This ensures K8s-managed resources (ALB, Target Groups) are cleaned up before VPC deletion.
Retained resources: Amazon EFS file systems use
removalPolicy: RETAINto protect tenant data. Aftercdk destroy, these remain in your account. To fully clean up:# List retained Amazon EFS (check for tenant data before deleting) aws efs describe-file-systems --query 'FileSystems[?contains(Name,`TenantEfs`)].FileSystemId' --output textDelete manually after confirming no data is needed.
After pulling new changes (git pull), update the deployed components:
# 1. Infrastructure + AWS Lambda code + Auth UI (AWS CDK deploys all three)
cd cdk && npx cdk deploy # Automatically detects existing stack
# 2. Re-apply platform manifests
bash scripts/deploy-platform.sh
# 3. Auth UI (if you use deploy-auth-ui.sh instead of AWS CDK BucketDeployment)
bash scripts/deploy-auth-ui.shNot all steps are needed for every update. Check the release notes for which components changed.
Contributions welcome. Please open an issue first to discuss changes.
Built with love on Amazon EKS + Amazon Bedrock
Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
SPDX-License-Identifier: MIT-0