Deploy OpenWorker from one YAML
Use one staging-deploy.yaml or prod-deploy.yaml for deployment choices and
advanced account/infrastructure overrides. There is no required setup-profile
file. Generated release manifests, plans, credentials and recovery journals are
separate installer artifacts, not additional configuration you must hand-author.
Complete Azure, AWS and local examples show the supported schema. The service installer remains experimental until the current release passes a fresh cloud install, agent/recovery and cleanup qualification.
owcluster deploy azure up --config staging-deploy.yaml --plan-only
owcluster deploy aws up --config prod-deploy.yaml --plan-only
Remove --plan-only to review/approve provisioning. --accept-plan DIGEST and
--accept-bootstrap DIGEST provide separate noninteractive approvals of the exact
saved review. The command never treats an uncertain result as permission to
recreate resources. Rerun with the original state to reconcile unfinished work.
Application choices
apiVersion: ows/v1alpha1
name: my-agents
cloud:
provider: azure
subscriptionId: "11111111-2222-4333-8444-555555555555"
region: westus2
nodes:
count: 2
instanceType: Standard_D4as_v7
managementUI:
enabled: true
port: 8444
observability:
enabled: true
retentionDays: 14
models:
anthropic: {env: ANTHROPIC_API_KEY}
advanced:
# Inline the provider/identity overrides shown in the complete example.
release: /absolute/path/to/generated-release/release.json
infrastructure: {}
azure:
resourceGroup: rg-my-agents
setupObjectId: "22222222-3333-4444-8555-666666666666"
setupPrincipalType: User
The snippet omits infrastructure detail for readability; copy a complete example.
Fixed node count or a CPU/memory size range is supported; autoscaling is roadmap.
models supplies env/file credential references, currently OpenAI and Anthropic.
It is not a general provider-routing configuration API. The installer stages keys
privately after approval. Completion reports modelSmokeTest: not_run; deploy and
submit an agent separately through the service SDK.
Advanced inputs and prerequisites
advanced.infrastructure embeds the reviewed provider inputs previously kept in
setup.json. AWS uses advanced.aws.operatorArn; Azure uses
advanced.azure.resourceGroup, setupObjectId and setupPrincipalType.
advanced.release selects a generated, checksum-verified release manifest with
pinned service/runner/UI images and chart/CA/OpenShell artifacts. Relative paths
in deployment YAML resolve against that YAML's directory. Generated release
manifest paths are absolute. Legacy explicit advanced.setupProfile import is
compatibility-only; new files do not use it or a hidden default account profile.
Install owcluster, the cloud CLI, Docker, Helm, kubectl and OpenSSL. Azure also needs
kubelogin and SSH for its optional management VM; AWS's SSM bridge needs the
Session Manager plugin. Sign in using your existing cloud identity. The installer
requires the reviewed provisioning/RBAC scopes, not merely a successful login.
Current Azure prerequisites still include existing workload/control resource groups, an Entra administrator group, versioned database password in an RBAC Key Vault with ARM deployment access, and an independent pinned cleanup image. AWS requires the reviewed installer identity and a versioned S3 cleanup package. The native approval/bootstrap flow handles its supported grants; it does not magically create all account prerequisites. Default-pin discovery, account setup wizard and publicly distributed release artifacts remain separate launch work.
Package the matching release using package.py --help: Linux ARM64 for AWS,
AMD64 for Azure, including service-ui when enabling the UI. New cloud deployments
require the reviewed OpenShell0.1.7 or0.1.8 archive. The current version does not assume
public service images or a published Helm chart are already available.
Cost estimates, not spending caps
There is no required budget field in new deployment YAML. Check an optional
estimated-cost ceiling in plan mode:
owcluster deploy azure up --config staging-deploy.yaml --plan-only \
--max-estimated-cost-usd 100
The CLI queries provider retail catalogs for the selected resources and saves a
separate cost-estimate.json with timestamp, sources, quantities, missing prices,
planned hours and ceiling result. The estimate covers modeled fixed infrastructure
plus an explicit allowance; model usage, variable traffic/storage/requests, taxes,
discounts and unmodeled resources are not a bill prediction or hard cap. A ceiling
check fails if modeled prices are incomplete or exceed the ceiling. Without a
ceiling, unknown prices remain visible rather than becoming zero or a passing
budget claim. The option requires --plan-only.
Quotes expire after30minutes. Rerun plan mode to refresh the quote for the same frozen resource intent; its review digest and original operations do not change. Before a new deployment is admitted, expired or failed quote checks require a fresh plan review. Already dispatched deployments retain their original approval for recovery; price expiry never causes replacement/recreation. Estimates are not continuously enforced during execution. Cloud billing alerts are separate. Earlier experimental YAML/native CLI budgeting is compatibility behavior only.
Optional trial expiry
Without trial, the workload does not automatically expire. For disposable data:
trial:
expiresAfterHours: 6
deleteData: true
This enables independent expiry cleanup and explicitly permits workload data deletion. It is not a guaranteed cost cap; failures and retained control/audit resources can still incur costs. Persistent destructive teardown remains refused pending a separately reviewed data-retention workflow.
UI, trace storage and private access
managementUI.port is the operator's loopback HTTPS preview port. The CLI derives
https://localhost:PORT; internal browser origin/CSRF checks remain strict. It
emits port-forward commands and CA information, not tokens in URLs. Use a current
service token to sign in. This does not create public ingress, DNS or a certificate
for a company domain. Existing ingress users can retain the legacy explicit origin;
owcluster management-ui enable --ingress-url https://manage.example.com selects it.
Do not combine an explicit ingress origin and preview port.
Use --keep-access to retain AWS SSM/Azure SSH private management connectivity.
The original Kubernetes endpoint/hostname/CA stay verified through the loopback
route. connection.json records API and enabled UI connection instructions.
After an installation has recorded its original private-access route,
--access-only --keep-access reopens that route for diagnosis without advancing
installation or replacing resources. Stop the command to close the route.
For Azure operators whose installation host cannot reach the database bootstrap
secret in Key Vault, --database-password-file /private/database-password can
supply that same database password. The file must be a bounded regular UTF-8 file
with private permissions (0600 on Unix). Its content is not printed or saved in
the installation review; the database host and TLS checks remain unchanged.
UI and recording are independent and disabled unless configured. Recording uses
its own connection/schema on execution PostgreSQL by default; an existing Secret
selected by observability.databaseSecret may provide database-url and ca.crt
for a separate database. That Secret must be staged in the owned namespace before
service installation. Enabling initializes its schema before startup.
owcluster management-ui enable --state ./deployment-state/installation --port 8444
owcluster management-ui disable --state ./deployment-state/installation
owcluster management-ui status --state ./deployment-state/installation
owcluster observability enable --state ./deployment-state/installation --retention-days 14
owcluster observability disable --state ./deployment-state/installation
owcluster observability status --state ./deployment-state/installation
The commands save feature overrides in the installer journal, retain the original review and verify cluster/namespace/Helm identities. Unknown upgrades are observed, never replayed. Disabling recording retains stored data; its retention worker is also stopped until recording resumes. Execution state, security audit and independent OTLP exporters remain separate. A UI rollout requires sign-in again.
For a separately signed API certificate, the UI can select a public CA Secret/key
through caSecret/caKey or --ca-secret/--ca-key. Without an override it trusts
the self-signed API certificate generated by the installer. The UI process gets
only public trust material; its TLS sidecar gets the separate private-key mount.