deer-flow/deploy/helm/deer-flow

DeerFlow Helm Chart

Deploys the full DeerFlow stack to Kubernetes: gateway (backend + embedded LangGraph runtime), frontend (Next.js), nginx (internal reverse proxy preserving the compose routing), and the provisioner (K8s-native sandbox that spawns code-execution Pods on demand).

This chart translates the production docker/docker-compose.yaml into native Kubernetes resources. No existing repo files are modified.

Prerequisites

• A Kubernetes cluster (Docker Desktop K8s, OrbStack, kind, k3d, or a real cluster).
• kubectl + helm 3.8+ installed (OCI registry support stabilized in 3.8; earlier 3.x needs HELM_EXPERIMENTAL_OCI=1).
• The three DeerFlow images — either the published ones (see "Install the published chart" below) or built locally (see step 1).
• An Ingress controller (e.g. ingress-nginx) if you enable ingress.

Install the published chart (GHCR)

The chart and all three images are published to GHCR on every v* release tag (see .github/workflows/container.yaml and chart.yaml). Skip the build step and install directly:

helm install deer-flow oci://ghcr.io/<owner>/charts/deer-flow \
  --version <version> \
  -n deer-flow --create-namespace \
  -f my-values.yaml

where <owner> is the GitHub owner the chart is published from and <version> matches the release tag without the leading v (tag v0.1.0 → --version 0.1.0).

Note: the helm chart is new in 2.1.0 - no chart was published before it. It publishes to oci://ghcr.io/<owner>/charts/deer-flow (the charts/ prefix keeps it distinct from the deer-flow-{backend,frontend,provisioner} image packages).

Point the chart at the published images:

image:
  registry: ghcr.io/<owner>     # owner prefix; images are <owner>/deer-flow-<name>
  tag: "<version>"              # match the release tag (sans leading `v`)
  pullSecrets:
    - { name: regcred }         # only if the GHCR package is private

The chart's gatewayImage / frontendImage / provisionerImage defaults already match the published image names (deer-flow-backend, deer-flow-frontend, deer-flow-provisioner), so only registry and tag are required. New GHCR packages default to private — flip the package to public in its GHCR settings page for unauthenticated pulls, otherwise create a pull secret (step 1) and reference it via image.pullSecrets.

The OCI chart and the images are versioned independently of the chart's appVersion; always set image.tag to the release that matches your chart --version unless you have a reason to pin differently.

1. Build & push images (custom builds only)

Skip this section if you're using the published chart above. To build the images yourself from the existing Dockerfiles:

REGISTRY=ghcr.io/yourorg
TAG=latest

# backend - build with the `postgres` extra so multi-replica deploys can use
# shared Postgres (matches the published image)
docker build -t $REGISTRY/deer-flow-backend:$TAG --build-arg UV_EXTRAS=postgres -f backend/Dockerfile .
# frontend
docker build -t $REGISTRY/deer-flow-frontend:$TAG -f frontend/Dockerfile .
# provisioner
docker build -t $REGISTRY/deer-flow-provisioner:$TAG -f docker/provisioner/Dockerfile docker/provisioner

docker push $REGISTRY/deer-flow-backend:$TAG
docker push $REGISTRY/deer-flow-frontend:$TAG
docker push $REGISTRY/deer-flow-provisioner:$TAG

These names match the chart's gatewayImage / frontendImage / provisionerImage defaults, so only image.registry and image.tag need to point at them.

If your registry needs auth, create a pull secret:

kubectl create secret docker-registry regcred \
  --docker-server=ghcr.io \
  --docker-username=youruser \
  --docker-password=yourtoken \
  -n deer-flow

2. Configure values

Copy and edit values.yaml → my-values.yaml. At minimum set:

image:
  registry: ghcr.io/yourorg
  tag: latest
  pullSecrets:
    - { name: regcred }

ingress:
  enabled: true
  className: nginx
  host: deer-flow.example.com
  tls:
    enabled: true
    secretName: deer-flow-tls

secrets:
  OPENAI_API_KEY: sk-...
  # add channel tokens, search keys, etc. as needed

Provide your model config under config (keep secrets as $VAR references — they resolve from the secrets map):

config: |
  config_version: 26
  models:
    - name: gpt-4
      use: langchain_openai:ChatOpenAI
      model: gpt-4
      api_key: $OPENAI_API_KEY
      request_timeout: 600.0
  sandbox:
    use: deerflow.community.aio_sandbox:AioSandboxProvider
    provisioner_url: http://provisioner:8002
  database:
    backend: postgres
    postgres_url: $DATABASE_URL
  checkpointer:
    type: postgres
    connection_string: $DATABASE_URL
  stream_bridge:
    type: redis   # cross-pod SSE; URL from DEER_FLOW_STREAM_BRIDGE_REDIS_URL
  # Tools MUST be listed explicitly - the agent gets none otherwise
  # (BUILTIN_TOOLS only adds present_file + ask_clarification). The chart
  # default in values.yaml enables the sandbox tools + web tools (web_search,
  # web_fetch, image_search - no API key); when you override `config:`, copy
  # them in. Full list in values.yaml / config.example.yaml. The web tools need
  # outbound egress from the gateway pod.
  tool_groups:
    - name: web
    - name: file:read
    - name: file:write
    - name: bash
  tools:
    - name: web_search
      group: web
      use: deerflow.community.ddg_search.tools:web_search_tool
      max_results: 5
    - name: web_fetch
      group: web
      use: deerflow.community.jina_ai.tools:web_fetch_tool
      timeout: 10
    - name: image_search
      group: web
      use: deerflow.community.image_search.tools:image_search_tool
      max_results: 5
    - name: bash
      group: bash
      use: deerflow.sandbox.tools:bash_tool
    # also: ls, read_file, glob, grep, write_file, str_replace (see values.yaml)

$DATABASE_URL is injected from the postgres Secret (see below). The checkpointer: section is required for multi-replica operation — the LangGraph Store (cross-thread memory + thread list) reads it and does not fall back to database:. stream_bridge.type: redis is the default and routes live SSE events through the bundled redis StatefulSet (or redis.external). Because config: is a single override blob, a partial config: replaces the chart default entirely - keep the tools:/tool_groups: block (or the agent will have no tools) and the sandbox:/database:/checkpointer:/stream_bridge: sections shown above.

3. Install (from a local chart checkout)

For a custom build or local development, install from the chart directory:

helm install deer-flow deploy/helm/deer-flow \
  -n deer-flow --create-namespace \
  -f my-values.yaml

4. Verify

kubectl -n deer-flow get pods
kubectl -n deer-flow port-forward svc/nginx 2026:2026
curl http://localhost:2026/health          # gateway health via nginx

Hit the Ingress host (map it in /etc/hosts for local clusters) to load the UI.

Provisioner sanity check:

kubectl -n deer-flow exec deploy/deer-flow-provisioner -- curl -s localhost:8002/health

Architecture notes

• PostgreSQL is the default database. A bundled single-instance postgres StatefulSet (postgresql.enabled: true) runs in the namespace and the gateway connects via the in-cluster Service. The DSN is auto-generated into a Secret (key database-url) and injected as DATABASE_URL; config.yaml references it as $DATABASE_URL in database.postgres_url. Schema is bootstrapped automatically on gateway startup (alembic create_all + stamp head). For real HA, disable the bundled instance and point at a managed DB:

  postgresql:
    enabled: false
    external:
      host: mydb.example.com   # or set databaseUrl / existingSecret
      port: 5432
      database: deerflow
      username: deerflow
      password: changeme
  

• Graceful shutdown & memory drain. The gateway pod sets terminationGracePeriodSeconds (default 45s, overridable via gateway.terminationGracePeriodSeconds) plus an optional preStop sleep (gateway.preStopSleepSeconds, default 5s). The grace period MUST exceed the Gateway's graceful-shutdown work — channel stop (~5s) plus the memory-queue drain (memory.shutdown_flush_timeout_seconds, default 30s) plus a buffer — because the drain runs on a daemon thread and K8s SIGKILLs anything still running at the end of the grace window. K8s defaults to 30s, which SIGKILLs the drain mid-flight and silently re-introduces the memory loss the drain is fixing. When you raise memory.shutdown_flush_timeout_seconds, raise gateway.terminationGracePeriodSeconds to match (channel stop + drain + buffer).
• Gateway replicas. Postgres + the Redis stream bridge together make the gateway's persisted state (checkpointer + run/thread metadata) and live stream path cross-pod-safe. The default is still 1 replica: do not raise gateway.replicas past 1 yet. Run control — create_or_reject dedup, cancel, and orphan reconciliation — is still worker-local (in-process asyncio.Lock + in-memory record.task), tracked by issue #3948. With >1 replica a double-submit can create two runs on one thread (checkpoint corruption), a cancel can land on a non-owner pod (409), and a crashed pod's runs stay pending/running forever. Stay on 1 replica until that work lands.
• Redis stream bridge. A bundled single-instance redis StatefulSet (redis.enabled: true, redis:7-alpine) runs in the namespace and the gateway connects via the in-cluster Service. Per-run SSE events are stored in Redis Streams (PR #3191) so a client connected to any gateway pod receives live events and reconnect resumes from Last-Event-ID. The URL is auto-generated into a Secret (key redis-url) and injected as DEER_FLOW_STREAM_BRIDGE_REDIS_URL; config.yaml sets stream_bridge.type: redis by default. No-auth by default (ClusterIP isolation, matching compose); set redis.auth.password to enable AUTH. For a managed Redis, disable the bundled instance and point at it via redis.external.
• Persistence. A PVC (<release>-home) backs /app/backend/.deer-flow (sqlite DB, memory, custom agents, per-thread user-data). The gateway mounts it with subPath: deer-flow so the layout matches the provisioner's PVC user-data mode. Default ReadWriteOnce; use ReadWriteMany (NFS) on multi-node clusters so sandbox Pods on other nodes can mount it.
• Provisioner RBAC. The provisioner gets a ServiceAccount with a namespaced Role (get/list/watch/create/delete on pods + services) and a narrow ClusterRole (namespace get/create). It uses in-cluster service-account creds — no kubeconfig mount. The unused update/patch/pods-exec/events verbs were dropped (audited against docker/provisioner/app.py).
• Skills. Disabled by default (emptyDir at /app/skills). Populate via skills.existingClaim or skills.configMap, or bake skills into a custom gateway image.

Security

Enforced posture

All workloads run as non-root with all Linux capabilities dropped. No container escalates privileges or runs as uid 0.

workload	runAsUser	fsGroup	writable-path handling
gateway	1000	1000	.deer-flow PVC group-writable via fsGroup; PYTHONDONTWRITEBYTECODE=1 suppresses .pyc writes; UV_CACHE_DIR=/tmp
frontend	1000 (node)	1000	emptyDir at /app/frontend/.next/cache (root-owned in the image)
nginx	101 (nginx)	101	command writes the rendered config to /tmp/nginx.conf and loads nginx -c /tmp/nginx.conf (since /etc/nginx is root-owned); emptyDir at /var/cache/nginx
provisioner	1000	—	no PVC; PYTHONDONTWRITEBYTECODE=1
postgres	999 (postgres)	999	official postgres:16 entrypoint detects non-root and skips the chown/gosu dance; data PVC group-writable via fsGroup
redis	999 (redis)	999	official redis:7-alpine entrypoint detects non-root and skips the gosu dance; data PVC group-writable via fsGroup

Every container sets:

• runAsNonRoot: true
• allowPrivilegeEscalation: false
• capabilities.drop: ["ALL"]
• seccompProfile: { type: RuntimeDefault }

All listening ports are >1024 (8001 / 3000 / 2026 / 8002 / 5432), so no NET_BIND_SERVICE capability is required.

ConfigMap rollout. ConfigMaps mount via subPath, which does not receive in-place updates — a helm upgrade that changes only a ConfigMap would leave pods on stale config. Each pod template carries a checksum/* annotation (SHA256 of the rendered ConfigMap): checksum/config + checksum/extensions on the gateway, checksum/nginx on nginx. Any content change alters the pod spec and triggers a rolling restart.

Resource defaults. Every workload ships with modest requests+limits in values.yaml; override per workload (gateway.resources, frontend.resources, nginx.resources, provisioner.resources, postgresql.primary.resources, redis.primary.resources).

Not yet enforced (deferred hardening)

These are intentionally not set in this chart revision. Each can be added per-workload with testing:

• readOnlyRootFilesystem: true — makes the container's root filesystem immutable so a compromised process can't persist changes to the image. Not enabled because it requires auditing every runtime write path and mounting an emptyDir over each. Known paths:
  • gateway / frontend / nginx / provisioner: /tmp (uv cache, python tempfiles, the nginx config + pid, node temp) — one emptyDir at /tmp each.
  • postgres: /tmp and /var/run/postgresql (the Unix-socket dir). The first four are mechanical. postgres is the hard case — the official image writes its socket to /var/run/postgresql and isn't designed for a read-only root, so it may need socket-path redirection (PGHOST/unix_socket_directories). Optionally, add USER directives to the backend/Dockerfile, frontend/Dockerfile, and docker/provisioner/Dockerfile so the images are non-root by default (defense in depth — the chart already forces the uid via securityContext, so this is not required). A cluster enforcing the restricted Pod Security Admission standard would require this setting.
• Provisioner RBAC narrowing. The Role grants get/list/watch/create/delete on pods and services in the namespace (update/patch/pods-exec/events were dropped as unused). These verbs still apply to all Pods in the namespace, not just sandbox Pods — RBAC can't scope by label, so the remaining options are a dedicated sandbox namespace or admission control (OPA/Kyverno).
• startupProbe. Workloads have readiness + liveness probes but no startup probe. The gateway's livenessProbe.initialDelaySeconds: 30 covers slow starts today; a startupProbe would let it take arbitrarily long to initialize without risking a liveness kill during a cold start (e.g. slow model config load).

None of these affect correctness of the current deployment.

Migrating an existing volume to non-root

fsGroup does not apply to subPath mounts, and it changes group ownership but not file mode — so a PVC written by an earlier root run (e.g. a cluster that ran the gateway as root before enabling this hardening, or a backup restore of root-owned files) will keep files like .jwt_secret at 0600 root:root. The non-root gateway (uid 1000) then can't read them and crashes on the first auth request with RuntimeError: Failed to read JWT secret from .../​.jwt_secret.

Fresh installs are unaffected — uid 1000 creates every file as 1000:1000.

To fix an existing root-written PVC, run a one-shot root pod that chowns the volume to the gateway uid (1000), then restart the gateway:

cat <<'EOF' | kubectl apply -n deer-flow -f -
apiVersion: v1
kind: Pod
metadata: { name: fix-home-perms, namespace: deer-flow }
spec:
  restartPolicy: Never
  containers:
    - name: chown
      image: busybox:1.36
      command: ["sh", "-c"]
      args: ["chown -R 1000:1000 /home-pvc/deer-flow && chmod -R g+rwX /home-pvc/deer-flow"]
      volumeMounts:
        - { name: home, mountPath: /home-pvc }
  volumes:
    - name: home
      persistentVolumeClaim: { claimName: deer-flow-deer-flow-home }
EOF
kubectl -n deer-flow wait --for=condition=Ready pod/fix-home-perms --timeout=30s
kubectl -n deer-flow delete pod fix-home-perms
kubectl -n deer-flow rollout restart deploy/deer-flow-deer-flow-gateway

(On a single-node cluster the fix pod can mount the RWO PVC concurrently with the gateway; on multi-node, scale the gateway to 0 first.) A durable alternative — an opt-in root volumePermissions initContainer that chowns on every start (the Bitnami pattern) — is not yet wired into this chart; it would introduce a root container, so it's left as an operator decision for now.

Sandbox Service type

The provisioner exposes each sandbox Pod behind a per-sandbox Service whose type is controlled by provisioner.sandboxServiceType (default ClusterIP).

ClusterIP (default, recommended). The provisioner returns a cluster-DNS URL - http://sandbox-<id>-svc.<namespace>.svc.cluster.local:8080 - so the gateway reaches its sandbox entirely inside the cluster network. No node IP, no high port, and the code-execution surface is not bound on every node's interfaces. This is correct for the chart, where the gateway always runs in-cluster.

NodePort (Docker-Compose/hybrid escape hatch). Set provisioner.sandboxServiceType: NodePort only when the gateway is not in K8s (e.g. the compose dev path, where the gateway is a container reaching sandbox Pods on the host's Docker Desktop K8s). The provisioner then returns http://{NODE_HOST}:{NodePort}. NODE_HOST defaults to the provisioner pod's node IP via the downward API (status.hostIP); because a NodePort is exposed on every node, the gateway can reach <node-IP>:<NodePort> on most clusters without configuration. Override provisioner.nodeHost only if your CNI or network policy blocks pod->node-IP traffic:

kubectl get nodes -o wide    # use INTERNAL-IP or EXTERNAL-IP

provisioner:
  sandboxServiceType: NodePort
  nodeHost: 192.168.x.x

On multi-node clusters, also switch persistence.home.accessMode to ReadWriteMany (this is orthogonal to the Service type - it governs whether a sandbox Pod can be scheduled on a node other than the gateway's).

Lint / dry-run

helm lint deploy/helm/deer-flow
helm template deer-flow deploy/helm/deer-flow -n deer-flow -f my-values.yaml | \
  kubectl apply --dry-run=client -f -

Uninstall

helm uninstall deer-flow -n deer-flow
# the PVC is NOT deleted by default — remove it manually if desired:
kubectl -n deer-flow delete pvc -l app.kubernetes.io/instance=deer-flow