Outbound connectivity reference

Every outbound destination the platform can generate, on one page, so a firewall or proxy allowlist can be written from this page alone rather than assembled from prose. The policy that governs this traffic, deny by default behind a hostname allowlist, is in Security assurance; how each route is set up is in Connectivity implementation. Two reading notes:

“Install-time” versus “runtime” is the column that matters for review. A self-hosted install needs the install-time rows only while installing or upgrading; at runtime the only Galtea-side traffic is the registry token refresh, and everything else goes to endpoints you own or chose. This is checkable: grep the values files of the package you receive for outbound hostnames, and the registry endpoint is the only Galtea-side address in them.
Hostnames written as <placeholder> are values specific to your deployment: Galtea issues the registry account and the <tenant> name, and the provider, product, storage and SMTP hosts are yours.

The tables

Where each destination is

Destination	Hostname	Port	Direction
Galtea image and chart registry (AWS ECR)	`<account>.dkr.ecr.eu-west-1.amazonaws.com`, account issued by Galtea	443	Cluster to internet
NetBird operator image and chart	`ghcr.io`	443	Cluster to internet
Redis-HA Helm chart	`dandydeveloper.github.io`	443	Cluster or CI to internet
Redis-HA container images	`public.ecr.aws` (Redis, Sentinel and HAProxy), `quay.io` (the metrics exporter) and `docker.io` (the init image)	443	Cluster to internet
VPN control plane, management	`vpn.platform.<tenant>.galtea.ai`	TCP 443	Your peers (devices, routing VMs, the operator) to internet, outbound only
VPN relay, fallback	`rels://relay-vpn.platform.<tenant>.galtea.ai`	TCP 443	Your peers to internet, outbound only
WireGuard data plane, direct path	The dedicated VPN load balancer endpoint Galtea exposes for your deployment, a stable public UDP address issued per deployment	UDP 51820	Your peers to internet, outbound only
LLM provider endpoints	Your provider account’s endpoint, for example your Azure OpenAI resource; routed solely through the LiteLLM gateway	443	Cluster (gateway) to provider
Your AI product endpoint	The hostnames you register for the egress allowlist, or a private address reached through the VPN	443, or your port over the VPN	Cluster (workers) to your endpoint
SMTP relay	Your relay’s host	Your relay’s port	Cluster (APIs) to relay
Object storage endpoint	Your bucket or account endpoint, for example `https://<your-storage-account>.blob.core.windows.net`	443	Cluster to storage, and your users’ browsers to storage, via signed URLs
Slack error webhook, opt-in and off by default	`hooks.slack.com`; the full URL, with its secret path, is issued when you enable the integration	443	Cluster to Slack

What each destination carries

Destination	What it carries	When it fires
Galtea image and chart registry (AWS ECR)	Container images and Helm charts, as OCI artifacts, pulled with read-only credentials	Install and upgrade; plus the token-refresh CronJob, which exchanges credentials every 2 hours so pulls keep working
NetBird operator image and chart	The NetBird Kubernetes operator, pulled from the NetBird project’s own registry	Install-time, and only if you use the Kubernetes way of joining the VPN
Redis-HA Helm chart	The unmodified public Redis-HA chart, from its maintainer rather than from Galtea’s registry	Install and upgrade
Redis-HA container images	The images the public Redis-HA chart references, pulled from their upstream registries	Install, upgrade, and whenever a pod is rescheduled onto a node without the image
VPN control plane, management	Identity, keys and route policy. Never application traffic	Runtime, continuous, in any deployment using the VPN
VPN relay, fallback	The still-encrypted WireGuard tunnel, when no direct peer-to-peer path exists. The relay cannot decrypt it	Runtime, only when a direct path fails
WireGuard data plane, direct path	Application traffic, encrypted end to end. The tunnel cannot be decrypted in transit	Runtime, the normal path. If it is blocked, traffic falls back to the relay above, costing latency, never confidentiality
LLM provider endpoints	Prompts and completions for evaluations and data generation	Runtime, during evaluation and generation
Your AI product endpoint	Test traffic against the system under test, with the credentials you supplied	Runtime, during test runs
SMTP relay	Transactional email: invitations, notifications	Runtime
Object storage endpoint	Uploaded files and stored artifacts	Runtime. The browser path is the one people forget: see Connectivity implementation
Slack error webhook, opt-in and off by default	Error events only: service name, error type and message, correlation identifiers. Never test data, prompts, model outputs or evaluation results	Runtime, only if you enabled it. Removing the URL from your values turns it off

What is deliberately absent

No telemetry endpoint. There is no analytics, usage or tracing destination pointing at Galtea. Tracing, if you enable it, targets a collector inside your own cluster. See Model 5: Self-hosted.
No license or activation callback. The deployment needs no online activation and no continuous connectivity to Galtea.
The database is not in these tables because it is reached inside your network, not over the internet: allow the cluster’s node subnet through its firewall per Prerequisites checklist.

Air-gapped installs remove the registry and chart rows by mirroring images and charts into your internal registry, by agreement: see Installation and lifecycle.

Overview

Deployment Models

Install and Operate

Connectivity and Security

Reference

Outbound connectivity reference

The tables

Where each destination is

What each destination carries

What is deliberately absent

​The tables

​Where each destination is

​What each destination carries

​What is deliberately absent

The tables

Where each destination is

What each destination carries

What is deliberately absent