Open-source AI coworkers that each get a computer of their own: a browser, files and tools, with every action decided before it happens and recorded after. Bring any AG-UI agent.
# EKS, with a computer for each person rather than one shared browser.
#
# Needs the agent-sandbox controller installed; CI renders it with --api-versions instead.
#
# CHECK THE CLUSTER HAS A DEFAULT STORAGECLASS BEFORE INSTALLING, because a fresh EKS cluster very
# often does not. `eksctl` creates `gp2`, which is not marked default and uses the in-tree
# `kubernetes.io/aws-ebs` provisioner that no longer exists in current Kubernetes. A volume asking
# for "the default" then never binds, the computer sits Pending, and nothing says why. Verified on a
# real 1.34 cluster; `kubectl get sc` shows it in one line.
#
# Either create a default class:
#
# provisioner: ebs.csi.aws.com, volumeBindingMode: WaitForFirstConsumer,
# annotated storageclass.kubernetes.io/is-default-class: "true"
#
# or name one here, which is what the line below does. `WaitForFirstConsumer` is not optional on any
# cloud: without it the volume is created in a zone picked before the pod is scheduled, and pods stick
# unschedulable with a node-affinity conflict, in multi-zone clusters only, so it passes every
# single-zone test.
config:
initialAdminEmails: [email protected]
intelligence:
apiUrl: https://api.intelligence.copilotkit.ai
gatewayWsUrl: wss://realtime.intelligence.copilotkit.ai
auth:
google:
clientId: example.apps.googleusercontent.com
publicUrl: https://openbot.example.com
postgresql:
enabled: false
database:
existingSecret: openbot-database
serviceAccount:
annotations:
eks.amazonaws.com/role-arn: arn:aws:iam::000000000000:role/openbot
externalSecrets:
enabled: true
secretStoreRef:
name: aws-secrets-manager
data:
- secretKey: key-encryption-key
remoteRef:
key: openbot/key-encryption-key
# Sessions are signed with this. The server refuses to start when sign-in is configured and it
# is missing, so every provider store has to carry it.
- secretKey: better-auth-secret
remoteRef:
key: openbot-better-auth-secret
- secretKey: intelligence-api-key
remoteRef:
key: openbot/intelligence-api-key
- secretKey: google-client-secret
remoteRef:
key: openbot/google-client-secret
- secretKey: computer-token
remoteRef:
key: openbot/computer-token
- secretKey: license-token
remoteRef:
key: openbot/license-token
ingress:
enabled: true
className: alb
annotations:
alb.ingress.kubernetes.io/scheme: internet-facing
alb.ingress.kubernetes.io/target-type: ip
server:
autoscaling:
enabled: true
minReplicas: 2
maxReplicas: 10
topologySpreadConstraints:
- maxSkew: 1
topologyKey: topology.kubernetes.io/zone
whenUnsatisfiable: ScheduleAnyway
labelSelector:
matchLabels:
app.kubernetes.io/name: openbot
app.kubernetes.io/component: server
computers:
# A computer for each person, which is the mode the rest of this chart exists for and the one nothing rendered
# until now. `shared` and `sandbox` produce different Deployments, different RBAC and a different
# pod template, so a target that only ever renders one of them checks half the chart.
mode: sandbox
persistence:
# Named here rather than in the chart, which is the whole point of a per-target values file: the
# chart must not name `gp3`, and a deployment on EKS should.
storageClass: gp3
# On, so these render in CI at all. They were written, reviewed and shipped without a single target
# producing one, which means nothing had ever checked they were valid YAML, let alone right. Off is
# still the chart's default: a policy on a cluster whose CNI ignores it does nothing, and on one that
# enforces it a wrong rule is an outage, so turning it on stays a deployment's decision.
networkPolicy:
enabled: true
# The only sandbox target, so the only one that renders the two policies carrying a rule for the
# Kubernetes API server — and the only one that has to name its service range. `172.20.0.0/16` is
# what EKS gives a cluster unless it was created with `--service-ipv4-cidr`. Named here rather than
# defaulted in the chart because the range is the cluster's and not the release's: the same line
# would be an outage on GKE. Leaving it out is now refused, which is the point of setting it here —
# the chart used to render a destination-less rule instead, permitting 443 and 6443 everywhere.
kubernetesApiCidr: 172.20.0.0/16
# This target's database is RDS rather than a pod, so egress to it has to be named. The chart
# refuses to render without this, which is how the omission was found: rendering the policies in CI
# at all is what exercised the refusal.
extraEgress:
- to:
- ipBlock:
cidr: 10.0.0.0/16
ports:
- port: 5432
protocol: TCP