hermes-operator

A Kubernetes operator (Go) for deploying and managing Nous Research Hermes Agent gateways declaratively via a HermesAgent Custom Resource.

Full design: docs/specification.md. Pinned upstream: Hermes Agent v2026.6.5 (vendored at third_party/hermes-agent).

What it does

One HermesAgent CR ⇒ one gateway = one Deployment(Recreate, replicas≤1) + one shared RWO PVC (sub-paths for data, ~/.local, Homebrew, and the dind store) + ConfigMaps + Service (+ optional Ingress) + an in-pod reloader sidecar. The gateway is a singleton (SQLite + gateway.lock), so the operator enforces replicas ∈ {0,1} and Recreate so the old pod releases the volume lock before the new one starts.

Highlights:

Renders config.yaml from typed CRD fields + deep-merged extraConfig.
Credentials via Secret refs — the operator never reads secret material (a Secret's resourceVersion feeds the config hash to trigger rollouts).
Single shared PVC with a settable size.
Model providers — typed model.providers[] for hermes built-ins (anthropic/openai/xai/openrouter/…, selected by name/alias) and custom OpenAI-compatible endpoints (with per-model context windows + key injection).
MCP servers — typed mcp.servers[] (stdio + http/sse) with per-server tool filtering and Secret-backed credentials injected for ${VAR} interpolation.
Bitwarden secrets — secrets.bitwarden syncs secrets via the bws machine account, with a Secret-backed access token and a custom/self-hosted serverURL.
Named hermes profiles — profile.name runs the agent under an isolated profiles/<name>/ home (own config/SOUL/skills/state) via HERMES_HOME.
Pre-install of optional deps onto the PVC (pip SDKs for channels/backends, honcho-ai, Apptainer for singularity) so they survive restarts.
Declarative skill activation and package installation — pip (init container, persisted on the PVC via PYTHONPATH) + Homebrew (no sudo).
Operator-managed Docker-in-Docker sidecar for the docker terminal backend, and an in-cluster kubeconfig written to ~/.kube/config.
A single shared Ingress config defaulted across every HTTP surface, with auto webhook ingress for webhook-capable channels.
Full podTemplate overlay strategic-merged over the operator's base pod, with operator invariants (shared-PVC mounts, config-hash, root-start) re-asserted.
Health-gated rollouts, finalizer-ordered teardown honoring reclaimPolicy.

Components (images → `harbor.bne1.ouchi.com.au/applications/`)

Component	Image	Build
Operator	`hermes-operator`	`images/operator/Dockerfile` (distroless)
Agent	`hermes-agent`	`images/agent/Dockerfile` (upstream + Homebrew + skill + s6 cont-init apt hook)
Reloader	`hermes-reloader`	`images/reloader/Dockerfile` (`FROM` agent + Go binary)

HermesAgent reference

apiVersion: hermes.nousresearch.io/v1alpha1, kind: HermesAgent. Everything below is under .spec. Optional fields show their default; — means no default (unset). Config-rendered fields note their config.yaml target as → key.

Top level

Field	Type	Default	Notes
`image`	string	required	Agent container image (the operator's derived image).
`imagePullPolicy`	string	`IfNotPresent`
`imagePullSecrets`	[]LocalObjectReference	—	Pull secrets (name refs).
`replicas`	int (0–1)	`1`	Singleton; `0` pauses (scales the Deployment to 0).
`presetRef.name`	string	—	A `HermesConfigPreset` deep-merged under this spec (CR wins).
`profile.name`	string	—	Run under a named hermes profile: config/SOUL/skills/state live in `/opt/data/profiles/<name>/` (HERMES_HOME points there). Pattern `^[a-z0-9][a-z0-9_-]{0,63}$`; reserved names rejected.
`soul`	string	—	Persona rendered to `$HERMES_HOME/SOUL.md` (the profile dir when `profile` is set).
`env`	[]corev1.EnvVar	—	Extra agent-container env (appended after operator vars).
`envFrom`	[]corev1.EnvFromSource	—	Mirrored onto the agent container.
`authJSONBootstrapSecretRef`	SecretKeyRef	—	Seeds `auth.json` once on first boot.
`extraConfig`	object (free-form)	—	Deep-merged into `config.yaml` (preserved verbatim).
`extraConfigPrecedence`	`merge`\|`override`	`merge`	`merge`: typed fields win; `override`: extraConfig wins.
`resources`	corev1.ResourceRequirements	—	Agent container resources.
`shmSize`	Quantity	—	`/dev/shm` emptyDir size (browser tools).
`hermesUID` / `hermesGID` / `fsGroup`	int64	`10000`	Run-as ids / fsGroup.
`runAsRoot`	bool	`true`	Governs the operator's init containers. The main agent container always starts as root: the image's s6-overlay `/init` requires root for cont-init bootstrap, then drops the hermes process to `hermesUID` via `s6-setuidgid`.
`podTemplate`	corev1.PodTemplateSpec	—	Strategic-merge overlay; operator invariants re-asserted after.

`model` (→ `model:`)

Field	Type	Default	Notes
`model.default`	string	—	Default model id, e.g. `claude-opus-4-7`. → `model.default`
`model.provider`	string	`auto`	Routing mode, or a `providers[]` entry name to make it active. → `model.provider`
`model.baseURL`	string	—	Endpoint override; defaults from the selected custom provider. → `model.base_url`
`model.apiMode`	string	—	Wire protocol: `chat_completions`, `anthropic_messages`, `codex_responses`, `bedrock_converse`. → `model.api_mode`
`model.contextLength`	int64	—	Global context window. → `model.context_length`
`model.maxTokens`	int64	—	Output cap. → `model.max_tokens`
`model.providers[]`	[]ProviderSpec	—	Built-in + custom providers (see below). Custom entries → `custom_providers:`.

model.providers[] — ProviderSpec

Field	Type	Default	Notes
`name`	string	required	Built-in name/alias (`anthropic`,`claude`,`openai`,`grok`,`xai`,`openrouter`,`zai`,`kimi`,`minimax`,`deepseek`,…) or a custom name.
`baseURL`	string	—	Set ⇒ custom OpenAI-compatible endpoint; omit ⇒ built-in (hermes knows the URL).
`apiMode`	string	—	As above.
`keyEnv`	string	—	Override the env var the key is injected under (else built-in's known var, or `<NAME>_API_KEY` for custom).
`keySecretRef`	SecretKeyRef	—	API key the operator injects as an env var. Omit for OAuth providers (`hermes login`) or keyless endpoints.
`models[]`	[]ProviderModelSpec	—	`{ name, contextLength }` — per-model context window (custom providers).

Selecting model.provider = a built-in alias resolves the key env automatically (claude→ANTHROPIC_API_KEY, grok→XAI_API_KEY, openai→OPENROUTER_API_KEY, …); for a custom active provider the operator derives model.base_url/api_mode from it.

`agent` (→ `agent:`)

Field	Type	Default	Notes
`agent.maxTurns`	int32	—	→ `agent.max_turns`
`agent.reasoningEffort`	`xhigh`\|`high`\|`medium`\|`low`\|`minimal`\|`none`	—	→ `agent.reasoning_effort`
`agent.disabledToolsets`	[]string	—	→ `agent.disabled_toolsets`

`compression` (→ `compression:`, hot-reloadable) · `memory` (→ `memory:`)

Field	Type	Default	Notes
`compression.enabled`	*bool	—
`compression.threshold`	string (number)	—	e.g. `"0.5"`
`compression.targetRatio`	string (number)	—	e.g. `"0.2"`
`compression.protectLastN` / `protectFirstN`	*int32	—
`memory.memoryEnabled`	*bool	—	→ `memory.memory_enabled`
`memory.userProfileEnabled`	*bool	—	→ `memory.user_profile_enabled`

`searxng` · `honcho`

Field	Type	Default	Notes
`searxng.url`	string	—	→ `SEARXNG_URL` env.
`honcho.baseURL`	string	—	→ `HONCHO_BASE_URL`. Honcho is "in use" when this or apiKeySecretRef is set.
`honcho.apiKeySecretRef`	SecretKeyRef	—	→ `HONCHO_API_KEY` (hosted Honcho).
`honcho.installPackage`	*bool	`true`	Pre-install `honcho-ai` onto the PVC (via the pip-install init).

`mcp` (→ `mcp_servers:`)

mcp.servers[] declares Model Context Protocol servers, rendered to config.yaml mcp_servers keyed by name. A server is stdio (set command) or http/sse (set url) — exactly one. Credentials go in secretEnv (a Secret key injected as a container env var) and are referenced with ${ENVNAME} inside headers/env values (hermes interpolates them at connect time; values pass through the operator untouched).

mcp.servers[] — MCPServerSpec

Field	Type	Default	Notes
`name`	string	required	Keys the server in `mcp_servers` (unique).
`enabled`	*bool	`true`	Toggle the server.
`transport`	`stdio`\|`http`\|`sse`	—	`stdio` implied by `command`; `url` implies `http` unless set to `sse`.
`command`	string	—	stdio: subprocess to launch. Exactly one of `command`/`url`.
`args`	[]string	—	stdio: command args.
`env`	map[string]string	—	stdio: subprocess env; values may use `${VAR}`.
`url`	string	—	http/sse: endpoint. Exactly one of `command`/`url`.
`headers`	map[string]string	—	http/sse: request headers; values may use `${VAR}` (e.g. `Bearer ${TOK}`).
`sslVerify`	*bool	`true`	http/sse: TLS verification.
`timeoutSeconds`	*int32	`120`	Per tool-call timeout.
`connectTimeoutSeconds`	*int32	`60`	Initial connection timeout.
`supportsParallelToolCalls`	*bool	`false`	stdio: allow concurrent tool calls.
`tools.include` / `tools.exclude`	[]string	—	Filter exposed tools (set at most one).
`secretEnv[]`	[]{`name`,`secretRef`}	—	Inject a Secret key as env var `name`; reference via `${name}`.
`extraConfig`	object	—	Deep-merged into this server (e.g. `oauth`, `sampling`); typed fields win.

`secrets` (→ `secrets:`)

secrets.bitwarden wires the agent to Bitwarden Secrets Manager (machine account + the bws CLI), rendered to config.yaml secrets.bitwarden. The machine-account access token is sensitive, so it is not rendered into config — accessTokenSecretRef is injected as the env var named by accessTokenEnv (default BWS_ACCESS_TOKEN) and hermes reads it at runtime. serverURL points bws at a custom region or self-hosted instance (e.g. https://vault.bitwarden.eu, or a Vaultwarden URL); leave it empty for the US cloud.

secrets.bitwarden — BitwardenSpec

Field	Type	Default	Notes
`enabled`	*bool	`false`	Master switch for Bitwarden sync. When `true`, `accessTokenSecretRef` is required.
`accessTokenSecretRef`	SecretKeyRef	—	Machine-account token; injected as env var `accessTokenEnv`.
`accessTokenEnv`	string	`BWS_ACCESS_TOKEN`	Name of the env var holding the access token.
`projectID`	string	—	Bitwarden project UUID to sync from.
`serverURL`	string	—	Custom/self-hosted endpoint (→ `server_url`). Empty = US cloud.
`cacheTTLSeconds`	*int32	`300`	In-process cache TTL for fetched secrets.
`overrideExisting`	*bool	`true`	Let Bitwarden values replace existing env vars (never the token var itself).
`autoInstall`	*bool	`true`	Let the agent download the `bws` binary on demand (needs egress).

`ingress` (shared, → Ingress objects)

Top-level spec.ingress is the shared Ingress config. When enabled (and dashboard.enabled), the dashboard is exposed at ingress.path. Its host/className/annotations/tls/pathType are inherited by apiServer and channel-webhook ingresses unless they set their own.

Field	Type	Default	Notes
`ingress.enabled`	bool	`false`
`ingress.host`	string	—	Base host; required (here or per-surface) when an ingress is enabled.
`ingress.className`	string	—	→ `spec.ingressClassName`.
`ingress.path`	string	`/`
`ingress.pathType`	`Exact`\|`Prefix`\|`ImplementationSpecific`	`Prefix`
`ingress.annotations`	map[string]string	—	Merged verbatim (nginx/cert-manager/etc.).
`ingress.tls[]`	[]{`hosts[]`,`secretName`}	—	TLS blocks.

`apiServer` (→ env, port 8642) · `dashboard` (→ env, port 9119)

Field	Type	Default	Notes
`apiServer.enabled`	bool	`false`	Requires `keySecretRef` + a non-localhost `host`.
`apiServer.host`	string	`0.0.0.0`
`apiServer.port`	int32	`8642`
`apiServer.keySecretRef`	SecretKeyRef	—	→ `API_SERVER_KEY`.
`apiServer.corsOrigins`	[]string	—
`apiServer.service`	ServiceSpec	—	Service type/annotations for this surface.
`apiServer.ingress`	IngressSpec	—	Per-surface ingress (inherits `spec.ingress`).
`dashboard.enabled`	bool	`false`
`dashboard.host`	string	`0.0.0.0`
`dashboard.port`	int32	`9119`
`dashboard.service`	ServiceSpec	—	(Dashboard ingress is the shared `spec.ingress`.)

`channels[]` (→ Service ports / Ingress / env)

Field	Type	Default	Notes
`type`	`telegram`\|`discord`\|`slack`\|`whatsapp`\|`signal`\|`email`\|`teams`\|`google_chat`	required
`enabled`	bool	`true`
`secretRef`	LocalObjectReference	—	Secret whose keys are the platform creds (injected via `envFrom`).
`webhookPort`	int32	`0`	`0` = polling; `>0` (or auto for webhook-capable channels) exposes a Service port.
`ingress`	IngressSpec	—	Opt into a webhook ingress (auto path `/webhooks/<type>`; only telegram serves an inbound webhook today).
`installDeps`	*bool	`true`	Pre-install the platform's pip deps (telegram/discord/slack).

`skills` (→ `skills:`)

Field	Type	Default	Notes
`skills.disabled`	[]string	—	Disable bundled skills.
`skills.platformDisabled`	map[string][]string	—	Per-platform disables.
`skills.externalDirs`	[]string	—	Read-only external skill dirs.
`skills.creationNudgeInterval`	int32	`15`
`skills.enablePackageManagementSkill`	*bool	`true`	Ensure the built-in package-management skill.
`skills.custom[]`	[]CustomSkill	—	`{ name, enabled=true,` one of `sourceRef{configMapName\|secretName} \| inline }`.

`packages` (→ init container + reloader)

Field	Type	Default	Notes
`packages.pip`	[]string	—	Installed to the PVC user-site (init container), loaded via `PYTHONPATH`.
`packages.pipImage`	string	`python:3.13-slim`	Must match the agent's Python (3.13 at the pinned tag).
`packages.brew`	[]string	—	Homebrew packages on the PVC (no sudo).
`packages.homebrewPrefix`	string	`/home/linuxbrew/.linuxbrew`

`runtime` (→ `terminal:`/`code_execution:`/`delegation:`)

Field	Type	Default	Notes
`runtime.terminalBackend`	`local`\|`docker`\|`ssh`\|`modal`\|`daytona`\|`singularity`	`local`	→ `terminal.backend`
`runtime.terminalTimeout`	int32	—	→ `terminal.timeout`
`runtime.installDeps`	*bool	`true`	Pre-install the backend's deps (modal/daytona SDKs, or Apptainer).
`runtime.codeExecution.timeout` / `.maxToolCalls`	int32	—	→ `code_execution.*`
`runtime.delegation.maxIterations` / `.maxConcurrentChildren`	int32	—	→ `delegation.*`
`runtime.docker.image`	string	`docker:27-dind`	DinD sidecar image.
`runtime.docker.rootless`	bool	`false`	Selects the `-dind-rootless` image when image is default.
`runtime.docker.socketTransport`	`unix`\|`tcp`	`unix`
`runtime.docker.mountCwdToWorkspace`	bool	`false`	→ `terminal.docker_mount_cwd_to_workspace`
`runtime.docker.resources`	ResourceRequirements	—	DinD sidecar resources.
`runtime.docker.storage.size`	Quantity	—	dind image store (subPath on the shared PVC).
`runtime.singularity.installImage`	string	`rockylinux:9`	Image for the Apptainer installer (needs curl/rpm2cpio/cpio).

`cron` (→ `/opt/data/cron/jobs.json`)

Field	Type	Default	Notes
`cron.pruneUnmanaged`	bool	`false`	Make the declared set authoritative.
`cron.jobs[]`	[]{`name`,`schedule`,`prompt`}	—	Seeded scheduled tasks.

`storage` · `serviceAccount` · `probes` · `kubeconfig`

Field	Type	Default	Notes
`storage.size`	Quantity	—	Shared PVC size. Mutually exclusive with `existingClaim`.
`storage.storageClassName`	string	—
`storage.existingClaim`	string	—	Use a pre-created PVC instead of provisioning.
`storage.reclaimPolicy`	`Retain`\|`Delete`	`Retain`	PVC fate on CR deletion.
`serviceAccount.create`	bool	`true`	Provision a per-agent SA + bind the reloader Role.
`serviceAccount.name`	string	— (agent name)	Required when `create=false`.
`serviceAccount.annotations`	map[string]string	—	Keyless cloud auth (IRSA/Workload Identity).
`serviceAccount.automountToken`	*bool	`true`
`probes.mode`	`auto`\|`exec`\|`http`	`auto`	`auto` = http when apiServer enabled, else exec. `http` requires `apiServer.enabled`.
`probes.liveness` / `probes.readiness`	ProbeSettings	—	`initialDelaySeconds`/`periodSeconds`/`timeoutSeconds`/`failureThreshold`/`successThreshold`.
`kubeconfig.enabled`	bool	`false`	Write `~/.kube/config` from the pod SA token. Requires `serviceAccount.automountToken != false`.

Validation (CEL)

storage.existingClaim is mutually exclusive with storage.size/storageClassName (set exactly one).
apiServer.enabled requires keySecretRef and a non-localhost host.
probes.mode=http requires apiServer.enabled.
ingress.enabled requires dashboard.enabled and ingress.host.
apiServer.ingress.enabled requires apiServer.enabled and a host (apiServer.ingress.host or spec.ingress.host).
channels[].ingress.enabled requires a host (channels[].ingress.host or spec.ingress.host).
each skills.custom[] must set exactly one of sourceRef or inline.
each mcp.servers[] must set exactly one of command (stdio) or url (http/sse); transport must match; tools sets at most one of include/exclude.
secrets.bitwarden.enabled requires accessTokenSecretRef.
profile.name must match ^[a-z0-9][a-z0-9_-]{0,63}$ and not be a reserved hermes name (hermes/default/test/tmp/root/sudo).
serviceAccount.name is required when serviceAccount.create=false.
kubeconfig.enabled requires serviceAccount.automountToken != false.

Examples

Minimal — a built-in provider (Claude)

apiVersion: hermes.nousresearch.io/v1alpha1
kind: HermesAgent
metadata:
  name: claude-bot
  namespace: agents
spec:
  image: harbor.bne1.ouchi.com.au/applications/hermes-agent:v2026.6.5
  imagePullSecrets: [{ name: harbor-pull }]
  storage: { size: 20Gi, storageClassName: fast-ssd }
  model:
    default: claude-opus-4-7
    provider: claude              # alias -> anthropic; key -> ANTHROPIC_API_KEY
    providers:
      - name: claude
        keySecretRef: { name: model-keys, key: anthropic }
  dashboard: { enabled: true }
  ingress:
    enabled: true
    host: claude-bot.example.com
    className: nginx
    tls: [{ hosts: [claude-bot.example.com], secretName: claude-bot-tls }]

Multiple providers, switchable

spec:
  model:
    default: claude-opus-4-7
    provider: claude              # active; /model can switch to the others
    providers:
      - { name: claude,  keySecretRef: { name: model-keys, key: anthropic } }
      - { name: grok,    keySecretRef: { name: model-keys, key: xai } }       # -> XAI_API_KEY
      - { name: openai,  keySecretRef: { name: model-keys, key: openrouter } } # -> OPENROUTER_API_KEY

Self-hosted OpenAI-compatible endpoint with per-model context

spec:
  model:
    default: gemma-4-e4b-it
    provider: llm-bne1            # selects the custom provider; base_url/api_mode derived
    providers:
      - name: llm-bne1
        baseURL: https://llm.bne1.ouchi.com.au/v1
        apiMode: chat_completions
        # keySecretRef omitted -> keyless endpoint (or seed auth.json)
        models:
          - { name: gemma-4-e4b-it,  contextLength: 131072 }
          - { name: qwen35-9b-opus,  contextLength: 65536 }

Docker terminal backend + in-cluster kubeconfig + brew tools

spec:
  runAsRoot: false
  serviceAccount: { create: false, name: agent-sa, automountToken: true }
  storage: { size: 50Gi, storageClassName: proxmox-local-ext4, reclaimPolicy: Retain }
  runtime:
    terminalBackend: docker        # operator injects a DinD sidecar on the shared PVC
  kubeconfig: { enabled: true }     # ~/.kube/config from the pod SA (kubectl works in-pod and in dind)
  packages:
    brew: [kubectl]                 # persisted on the PVC, also bind-mounted into dind

Messaging channels (telegram webhook + discord polling)

spec:
  ingress: { enabled: true, host: bot.example.com }   # shared host
  dashboard: { enabled: true }
  channels:
    - type: telegram
      secretRef: { name: bot-channels }   # must include TELEGRAM_WEBHOOK_SECRET for webhook mode
      ingress: { enabled: true }          # -> https://bot.example.com/webhooks/telegram (auto port)
    - type: discord
      secretRef: { name: bot-channels }   # outbound gateway websocket; webhookPort 0 (polling)

OpenAI-compatible API server + extraConfig escape hatch

spec:
  apiServer:
    enabled: true
    host: 0.0.0.0
    corsOrigins: ["*"]
    keySecretRef: { name: agent-secret, key: API_SERVER_KEY }
    ingress: { enabled: true }            # inherits spec.ingress host/tls
  ingress: { enabled: true, host: api.example.com }
  # Anything not modeled as a typed field can be merged into config.yaml:
  extraConfig:
    fallback_model: { provider: openrouter, model: anthropic/claude-sonnet-4 }
    auxiliary:
      compression: { provider: openrouter, model: google/gemini-3-flash }

MCP servers (stdio + http with a secret token)

spec:
  mcp:
    servers:
      # stdio: launched as a subprocess; token interpolated into its env.
      - name: github
        command: npx
        args: ["-y", "@modelcontextprotocol/server-github"]
        env:
          GITHUB_PERSONAL_ACCESS_TOKEN: ${GH_TOKEN}
        secretEnv:
          - name: GH_TOKEN
            secretRef: { name: mcp-secrets, key: github-token }
        tools:
          include: [create_issue, get_issue]   # narrow the exposed tools
      # http/sse: remote endpoint with a bearer token from a Secret.
      - name: remote
        transport: sse
        url: https://mcp.example.com/sse
        headers:
          Authorization: Bearer ${REMOTE_TOKEN}
        secretEnv:
          - name: REMOTE_TOKEN
            secretRef: { name: mcp-secrets, key: remote-token }
        extraConfig:                            # long-tail knobs (oauth, sampling)
          sampling: { enabled: true, max_rpm: 10 }

Bitwarden secrets (self-hosted server)

spec:
  secrets:
    bitwarden:
      enabled: true
      serverURL: https://vault.example.com   # custom region / self-hosted / Vaultwarden
      projectID: 11111111-2222-3333-4444-555555555555
      # The token rides in as env BWS_ACCESS_TOKEN (never written to config.yaml).
      accessTokenSecretRef: { name: bw-creds, key: token }

Run under a named hermes profile

spec:
  profile:
    name: staging          # HERMES_HOME=/opt/data/profiles/staging; config/SOUL/
                           # skills/memory/sessions isolated there. Gateway runs
                           # under `staging`; the default home is untouched.

A complete, production CR (custom endpoint + docker + kubeconfig + channels + searxng + honcho + dashboard ingress) lives at k8s-cluster/ouchi/bne1-cluster1/hermes/lana-hermesagent.yaml.

Layout

api/v1alpha1/                 # CRD types + CEL validation + deepcopy
internal/config/              # config.yaml renderer + config hash
internal/resources/           # child-object builders (pvc/cm/svc/ingress/sa/deployment)
internal/controller/          # reconciler + finalizer
cmd/operator, cmd/reloader/   # the two binaries
images/{operator,agent,reloader}/
config/                       # kustomize (crd, rbac, manager, samples)
charts/hermes-operator/       # Helm chart (CRDs + RBAC + manager)
third_party/hermes-agent/     # upstream submodule @ v2026.6.5

Develop

make generate manifests       # codegen + CRD/RBAC manifests
make build                    # build operator + reloader
make test                     # unit + envtest integration
make helm-lint helm-template  # render/validate the chart
kustomize build config/default

# Images (single-arch local, or buildx multi-arch + push to Harbor)
make docker-build
make docker-buildx IMG_REGISTRY=harbor.bne1.ouchi.com.au/applications VERSION=v0.1.0
make helm-push

Name		Name	Last commit message	Last commit date
.devcontainer		.devcontainer
.github/workflows		.github/workflows
api/v1alpha1		api/v1alpha1
charts/hermes-operator		charts/hermes-operator
cmd		cmd
config		config
dist		dist
docs		docs
hack		hack
images		images
internal		internal
test		test
third_party		third_party
.custom-gcl.yml		.custom-gcl.yml
.dockerignore		.dockerignore
.gitignore		.gitignore
.gitmodules		.gitmodules
.golangci.yml		.golangci.yml
AGENTS.md		AGENTS.md
CLAUDE.md		CLAUDE.md
Makefile		Makefile
PROJECT		PROJECT
README.md		README.md
go.mod		go.mod
go.sum		go.sum

Navigation Menu

Uh oh!

Repository files navigation

hermes-operator

What it does

Components (images → `harbor.bne1.ouchi.com.au/applications/`)

HermesAgent reference

Top level

`model` (→ `model:`)

`agent` (→ `agent:`)

`compression` (→ `compression:`, hot-reloadable) · `memory` (→ `memory:`)

`searxng` · `honcho`

`mcp` (→ `mcp_servers:`)

`secrets` (→ `secrets:`)

`ingress` (shared, → Ingress objects)

`apiServer` (→ env, port 8642) · `dashboard` (→ env, port 9119)

`channels[]` (→ Service ports / Ingress / env)

`skills` (→ `skills:`)

`packages` (→ init container + reloader)

`runtime` (→ `terminal:`/`code_execution:`/`delegation:`)

`cron` (→ `/opt/data/cron/jobs.json`)

`storage` · `serviceAccount` · `probes` · `kubeconfig`

Validation (CEL)

Examples

Minimal — a built-in provider (Claude)

Multiple providers, switchable

Self-hosted OpenAI-compatible endpoint with per-model context

Docker terminal backend + in-cluster kubeconfig + brew tools

Messaging channels (telegram webhook + discord polling)

OpenAI-compatible API server + extraConfig escape hatch

MCP servers (stdio + http with a secret token)

Bitwarden secrets (self-hosted server)

Run under a named hermes profile

Layout

Develop

About

Releases

Packages

Contributors

Languages

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

hermes-operator

What it does

Components (images → harbor.bne1.ouchi.com.au/applications/)

HermesAgent reference

Top level

model (→ model:)

agent (→ agent:)

compression (→ compression:, hot-reloadable) · memory (→ memory:)

searxng · honcho

mcp (→ mcp_servers:)

secrets (→ secrets:)

ingress (shared, → Ingress objects)

apiServer (→ env, port 8642) · dashboard (→ env, port 9119)

channels[] (→ Service ports / Ingress / env)

skills (→ skills:)

packages (→ init container + reloader)

runtime (→ terminal:/code_execution:/delegation:)

cron (→ /opt/data/cron/jobs.json)

storage · serviceAccount · probes · kubeconfig

Validation (CEL)

Examples

Minimal — a built-in provider (Claude)

Multiple providers, switchable

Self-hosted OpenAI-compatible endpoint with per-model context

Docker terminal backend + in-cluster kubeconfig + brew tools

Messaging channels (telegram webhook + discord polling)

OpenAI-compatible API server + extraConfig escape hatch

MCP servers (stdio + http with a secret token)

Bitwarden secrets (self-hosted server)

Run under a named hermes profile

Layout

Develop

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Components (images → `harbor.bne1.ouchi.com.au/applications/`)

`model` (→ `model:`)

`agent` (→ `agent:`)

`compression` (→ `compression:`, hot-reloadable) · `memory` (→ `memory:`)

`searxng` · `honcho`

`mcp` (→ `mcp_servers:`)

`secrets` (→ `secrets:`)

`ingress` (shared, → Ingress objects)

`apiServer` (→ env, port 8642) · `dashboard` (→ env, port 9119)

`channels[]` (→ Service ports / Ingress / env)

`skills` (→ `skills:`)

`packages` (→ init container + reloader)

`runtime` (→ `terminal:`/`code_execution:`/`delegation:`)

`cron` (→ `/opt/data/cron/jobs.json`)

`storage` · `serviceAccount` · `probes` · `kubeconfig`