Skip to content

README.md

Latest commit

 

History

History
261 lines (204 loc) · 9.8 KB

README.md

File metadata and controls

261 lines (204 loc) · 9.8 KB
Raw

OpenHands (Agent Canvas) on Kubernetes

This folder contains everything needed to deploy Agent Canvas — the self-hosted, single-tenant OpenHands deployment (web UI + agent-server + automations in one all-in-one image) — on Kubernetes, protected in front by oauth2-proxy authenticating against a GitHub organization.

This guide is intentionally cloud-agnostic: it works on any Kubernetes cluster (managed or not) that has an nginx ingress controller, cert-manager and a block StorageClass. Cluster-specific values are shown as examples to adapt, not to copy blindly.

Architecture

Internet
   │  https://agent.logora.com
   ▼
nginx Ingress (TLS via cert-manager)
   │  auth_url / auth_signin (auth_request)
   ▼
oauth2-proxy ── GitHub OAuth (org "Logora") → only org members pass
   │
   ▼
Agent Canvas (StatefulSet, PVC for persistent state)
  • Agent Canvas: official Helm chart, deployed as a StatefulSet (UI + agent-server + automations) with one PersistentVolumeClaim.
  • oauth2-proxy: GitHub OAuth gateway; only members of the GitHub org Logora can reach the UI.
  • Dedicated node: both workloads are pinned to a dedicated, tainted node so agent code execution does not compete with (or get evicted by) your other applications.

Prerequisites

  • Kubernetes 1.24+
  • Helm 3.x and kubectl configured against the target cluster
  • nginx ingress controller installed
  • cert-manager installed (automatic TLS)
  • A ReadWriteOnce block StorageClass for the PVC (update persistence.storageClassName in values.yaml — OVH example: csi-cinder-high-speed)
  • A GitHub OAuth App (see below)
  • An OpenRouter API key

Setup

1. GitHub OAuth App

  1. GitHub → Settings → Developer settings → OAuth Apps → New OAuth App.
  2. Set:
    • Homepage URL: https://agent.logora.com
    • Authorization callback URL: https://agent.logora.com/oauth2/callback
  3. Copy the Client ID and Client Secret into your local .env (see .env.example).
  4. Membership is restricted to the Logora org via --github-org=Logora in oauth2-proxy/deployment.yaml. Adjust if you want a different org or team.

2. Namespace and secrets

kubectl create namespace openhands

Create the secrets from a .env file (copy .env.example first and fill it in):

# LLM credentials (OpenRouter)
kubectl create secret generic openhands-llm \
  --namespace openhands --from-env-file=.env

# GitHub OAuth credentials for oauth2-proxy
kubectl -n openhands create secret generic openhands-github-oauth \
  --from-literal=OAUTH2_PROXY_CLIENT_ID="$(grep OAUTH2_PROXY_CLIENT_ID .env | cut -d= -f2)" \
  --from-literal=OAUTH2_PROXY_CLIENT_SECRET="$(grep OAUTH2_PROXY_CLIENT_SECRET .env | cut -d= -f2)" \
  --from-literal=OAUTH2_PROXY_COOKIE_SECRET="$(grep OAUTH2_PROXY_COOKIE_SECRET .env | cut -d= -f2)"

All secrets can instead be created out of band (Vault, SealedSecrets, your GitOps operator) and simply referenced by name — the manifests only read them via secretKeyRef.

3. Dedicated node

Pin everything to a dedicated node so agent workloads are isolated:

# Label + taint the target node (replace <node-name>)
kubectl label node <node-name> node-role.openhands.infra=true
kubectl taint nodes <node-name> openhands=true:NoSchedule

Persistence caveat: a taint/label applied with kubectl is lost if the node is later replaced by your provider (maintenance, resize, autoscaling). To make it survive, define the taint/label at the node pool template level if your provider supports it (e.g. OVH template.metadata on the node pool, or your cloud's equivalent). A node reboot alone does not lose a kubectl-applied taint (it lives in etcd).

Keep min_nodes = max_nodes on the pool if you don't want the node scaled down/replaced unnecessarily.

4. Install Agent Canvas

There is no published Helm repo — the chart lives in the OpenHands source tree. Clone it and install from the local path:

git clone https://github.com/OpenHands/OpenHands.git
cd OpenHands

helm install agent-canvas ./helm/agent-canvas \
  --namespace openhands \
  --values /path/to/openhands/values.yaml

Upgrade later with:

helm upgrade agent-canvas ./helm/agent-canvas \
  --namespace openhands --values /path/to/openhands/values.yaml

5. Deploy oauth2-proxy

kubectl apply -f oauth2-proxy/service.yaml
kubectl apply -f oauth2-proxy/deployment.yaml
kubectl apply -f oauth2-proxy/ingress.yaml

Verification

kubectl -n openhands rollout status statefulset/agent-canvas
kubectl -n openhands get pvc,pod,svc,ingress
kubectl -n openhands get pods -l app=openhands-oauth2-proxy

# Confirm the dedicated node scheduling
kubectl -n openhands get pods -o wide

Browse to https://agent.logora.com — you should be redirected to GitHub login and only members of the Logora org get through.

LLM model

The LLM is configured in the Agent Canvas UI (Settings → LLM), which stores it durably on the PVC (~/.openhands) and survives restarts/upgrades.

Suggested starter (OpenRouter):

  • Provider: OpenRouter (OpenAI-compatible / LiteLLM)
  • Base URL: https://openrouter.ai/api/v1
  • Model: openrouter/deepseek/deepseek-v4-flash-0731 (OpenRouter model ID for DeepSeek v4 Flash)
  • API key: an OpenRouter key sk-or-…

For heavier agentic coding you can switch to a stronger model, e.g. openrouter/anthropic/claude-opus-5 (Claude Opus 5), in the UI at any time. OpenRouter exposes every model under the openrouter/<id> prefix.

To configure purely via environment instead of the UI, uncomment the config.extraEnv block at the top of values.yaml (it references the openhands-llm secret) and run the helm upgrade above.

GitHub integration & PR automation

GitHub token (required)

OpenHands (OSS) authenticates to GitHub via a Personal Access Token. Add it in the UI: Settings → Integrations → GitHub Token.

  1. Create a fine-grained PAT (Developer settings → Personal access tokens → Fine-grained tokens), owned by the Logora org.
  2. Scope it to the repositories you want.
  3. Permissions (all "Read and write" unless noted):
    • Contents (create branches / commits / push)
    • Pull requests (open/manage PRs)
    • Issues (comment on tickets)
    • Metadata (read-only, mandatory)
  4. If the org enforces it, click Enable SSO next to the org.
  5. Paste the token in Settings → Integrations → GitHub Token and save.

OpenHands exports this as GITHUB_TOKEN into the agent shell, so git push, gh pr, comments and PRs all work from the workspace.

Auto-PR on @mention (issues / PRs)

Trigger: when the bot is @mentioned in a comment on an issue or a pull-request, the agent works on it and opens a PR. The most robust self-contained approach for OSS is a repository GitHub Actions workflow (runs where the code lives, no need to expose a webhook to Agent Canvas).

A ready-to-adapt template is provided in github-action/auto-pr-on-mention.yml. To enable it in a repository:

  1. Copy the file to <repo>/.github/workflows/auto-pr-on-mention.yml.
  2. Make sure the workflow's @mention (e.g. @openhands / /openhands) matches how your team summons the agent.
  3. Complete the agent invocation step in the workflow (the placeholder) to actually run the agent against the referenced issue/PR and open a PR.
  4. If the repository uses a service account bot to push, create a token with repo scope and store it as an Actions secret instead of relying on the default GITHUB_TOKEN.

The token from Settings → Integrations → GitHub Token is what the agent itself uses inside Agent Canvas to push/PR from its conversations.

Uninstall

# Remove the Helm release (PVC is retained by StatefulSet — delete explicitly to reset)
helm uninstall agent-canvas -n openhands
kubectl -n openhands delete pvc -l app.kubernetes.io/instance=agent-canvas

# Remove oauth2-proxy
kubectl delete -f oauth2-proxy/ingress.yaml
kubectl delete -f oauth2-proxy/deployment.yaml
kubectl delete -f oauth2-proxy/service.yaml

# Delete secrets + namespace
kubectl -n openhands delete secret openhands-llm openhands-github-oauth
kubectl delete namespace openhands

# Remove the node taint/label if no longer needed
kubectl taint nodes <node-name> openhands=true:NoSchedule-
kubectl label node <node-name> node-role.openhands.infra-

Troubleshooting

  • WebSocket disconnects / UI drops mid-turn: the ingress closes idle streams. Ensure the read/send timeout annotations (proxy-read-timeout, proxy-send-timeout) are set to at least 3600.
  • Auth loop / 401: check the GitHub OAuth App callback URL matches exactly, and that the requesting user is a member of the Logora org. Check oauth2-proxy logs: kubectl -n openhands logs -l app=openhands-oauth2-proxy.
  • Pod stuck Pending: kubectl -n openhands describe pod <pod> — usually a missing/unschedulable StorageClass or the dedicated node not having the expected taint/label (see the persistence caveat above).
  • OOMKill / restarts: check kubectl -n openhands top pods and increase resources.limits.memory in values.yaml.

Security notes

  • Agent Canvas (open-source) is unauthenticated, single-tenant: all agents share one pod/filesystem. oauth2-proxy provides the access control, but the pod itself is still shared state — treat the namespace as trusted.
  • The agent server can read/write the pod filesystem and run shell commands. Do not grant rbac unless you explicitly want the agent to mutate the cluster, and scope it to specific namespaces only.
  • OpenHands' own "native GitHub auth" (SSO/RBAC, per-user isolation) is a commercial (Enterprise) feature, not part of this open-source chart.