What does a data plane connect to?
Image pulls are made by your cluster’s nodes, not by the plane’s pods. Allow your image registry
(TCP 443) in the node or cloud firewall; the plane’s own egress restriction does not need it.
Query engines your plane starts are not covered by the plane’s egress restriction. They reach your
data sources directly and have their own outbound settings.
How do I restrict the plane’s outbound connections?
The data plane install has an optional profile,networkPolicy.egressAllowlist, that refuses every
outbound connection from the plane’s pods except the destinations above. It is off by default. Your
cluster must enforce network policies (Calico, Cilium, or your cloud’s equivalent); on a cluster that
does not, the rules have no effect.
Check two things before you turn it on:
- Cilium, including GKE Dataplane V2. A stock Cilium install does not match address ranges against
addresses inside the cluster, and your API server’s endpoint addresses are inside it. Without
Cilium’s
--policy-cidr-match-mode=nodes, the API server rule has no effect, the plane cannot reach its API server, and its worker stops at start. Calico and Antrea need no extra setting. - A namespace of its own. The profile restricts every Dippr pod in the plane’s namespace. Do not turn it on where the control plane runs in the same namespace (a combined install), or the control plane loses its outbound connections too.
my-plane.yaml
login.microsoftonline.com, TCP 443). Add one more
extra entry, named entra-id. Without it, an install that restricts egress and runs agents on Azure is refused:
extra.
Without the token service, the plane cannot sign in to your object store.
Every entry takes a
ports list to override the default. Pods in the plane’s own namespace, and data
apps in the apps namespace, can always reach each other.
For the API server, list its endpoint addresses, not only the service address: most clusters apply
the rules after translating the service address. Find them with
kubectl get endpointslices -n default -l kubernetes.io/service-name=kubernetes.
What happens if a value is missing?
The install refuses to start and names the setting, for example:0.0.0.0/0 or ::/0 (that would allow every address), an entry with no
ports, and anything that is not an address range, such as a host name.
How do I know it is working?
After you turn the profile on, the plane should still show Connected in Settings → Data planes. If it shows Disconnected, an address range is missing or wrong: check the control plane and job service ranges first. See Data plane status. Cloud services such as object storage, Snowflake or BigQuery answer from large, changing address ranges. If you cannot list them, use a private endpoint for the service, or leave the profile off and restrict outbound traffic at your cloud firewall instead.Can the plane’s MCP endpoint stay on a private network?
Yes. A plane whose MCP endpoint (<slug>.mcp.<plane domain>, and the shared mcp.<plane domain> on planes that have
one) resolves only to private addresses is supported. The plane still dials out to your control plane as above, so it
works and shows Connected. What changes is which AI clients can reach its MCP endpoint:
A private DNS record alone does not keep the endpoint private. The plane’s hostnames share one front
door, and it routes by hostname, so if any hostname of the plane is served through a public front door, an
internet caller can ask it for the MCP hostname. Keep the front door itself on the private network, or
restrict the addresses it accepts connections from. Which setting does that depends on which door your
plane uses; in the data plane chart:
- The cluster Ingress (
ingress.enabled, on by default): an allow-list annotation underingress.annotations, which is empty by default. - The chart’s load-balancer front door (
frontDoor.loadBalancer.enabled, off by default):frontDoor.loadBalancer.sourceRanges, which defaults to0.0.0.0/0. It has no effect while that door is off. - A Gateway API route (
frontDoor.gateway.enabled, off by default): the chart does not create the Gateway, so restrict who can reach the Gateway you run.
To sign in with OAuth, a local agent also has to reach your region’s account address (
https://app.<domain>, or
your deployment’s own account address) for mcp.<plane domain>, or your tenant’s address
(https://<slug>.app.<domain>) for <slug>.mcp.<plane domain>. When the plane’s certificate comes
from a private CA, it also has to trust that CA (see Certificates). An
agent that uses an API key on <slug>.mcp.<plane domain> does not sign in, so it needs only the MCP
endpoint (see Connect an agent).
Claude’s hosted connectors, the ones in the table’s second row, connect from Anthropic’s servers, not
from your machine, so a connector cannot reach an endpoint that only your network can. Claude Code is
not a hosted connector: it connects from the machine it runs on. To use a hosted connector, the MCP
endpoint has to be reachable from the internet. For Claude, Anthropic’s
custom connector guide
says to allow Anthropic’s IP addresses in your firewall.
Your query engines, data apps and app registry are reached the same way: only from machines that can
reach the plane’s addresses.