Skip to content

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.

4.7 KB
Raw Blame
# 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