Skip to navigation

Network and egress

This page applies to the Enterprise plan, where you run the data plane in your own VPC. It lists every network path the data plane’s components use, so you can pre-authorize firewall and security group rules before you install. The data plane only makes outbound connections to Postman. The control plane never connects in to your VPC.

With workload attestation, the control plane also reads your cluster’s public signing keys from its OIDC issuer URL, unless you paste the keys when you register the trust anchor. If your issuer isn’t reachable from the internet, select Paste JWKS. See Connect to the control plane.

Data flow

  1. Your clients send requests to the gateway Service on port 8080.
  2. The gateway pods keep an mTLS connection open to the control plane’s sync address for enrollment, certificate renewal, and configuration sync. If the pods can’t reach it, they never become ready.
  3. The gateway pods call your LLM providers, MCP servers, and APIs. With the MySQL secret store, they also read and write credentials in MySQL.
  4. The gateway calls the embedder inside the cluster for MCP tool search. The embedder downloads its model from huggingface.co at startup unless you run it offline.
  5. The collector sidecar in each gateway pod exports traces, metrics, and logs to your telemetry backend. With otelCollector.enabled: false, there’s no sidecar, so set OTEL_EXPORTER_OTLP_ENDPOINT to your backend and the gateway pods export directly. Otherwise telemetry is dropped.

Required egress

Allow these outbound paths from the cluster:

FromToPortPurposeRequired
Gateway podsControl plane sync addressThe port in the address Postman gives you, 9090 in the chart’s examples (gRPC over mTLS)Enrollment, certificate renewal, and configuration syncYes
Gateway podsYour LLM providers, MCP servers, and APIsEach upstream’s port, usually 443Request trafficYes
Gateway podsYour identity providers (OAuth and OIDC discovery, token, and JWKS URLs)Usually 443Client authentication and account connectionsIf you use them
Gateway pods and migration JobMySQL3306Secret store and schema migrationsWith the MySQL store
Collector sidecar, or gateway pods with the collector offYour telemetry backendYour backend’s port (the chart’s ClickHouse example uses 9440)Traces, metrics, and logsIf you export telemetry
Embedderhuggingface.co443Downloads the embedding model at startupUnless you run it offline
Gateway podsYour coordination RedisYour Redis portShared rate-limit countersOnly if you turn it on

To send outbound calls through a proxy, set global.env.HTTP_PROXY and HTTPS_PROXY. They apply to the gateway, collector, embedder, and migration Job.

On AWS, these paths map to security group egress rules on the nodes or pods that run the gateway and embedder. If the gateway’s ServiceAccount uses an IAM role (IRSA), the pods also need to reach the regional STS endpoint, or a VPC endpoint for STS. Without it, the credential exchange fails even when every other rule is correct.

If a proxy or load balancer between the data plane and the control plane ends TLS, the control plane can’t see the data plane’s client certificate, and certificate renewal fails. Let the sync connection pass through with TLS intact. If you run Istio, set istio.enabled so the sidecar doesn’t intercept the ports in istio.excludeOutboundPorts, 4317,4318,9090 by default. Only those ports are excluded, so if your sync address uses a port other than 9090, add it to that list.

Calls to private hosts

To protect against server-side request forgery, the gateway never connects to link-local, cloud metadata, or multicast addresses. It refuses loopback addresses by default. To let routes reach an upstream on loopback, set FABRIC_GW_SSRF_ALLOW_LOOPBACK_UPSTREAMS to "true". Private addresses (RFC 1918, carrier-grade NAT, and IPv6 unique local) are a different case for each kind of call:

  • Your upstreams. Routes can reach LLM providers, MCP servers, and APIs on private addresses by default. To block private upstreams, set FABRIC_GW_SSRF_BLOCK_PRIVATE_UPSTREAMS to "true". Then list any private hosts your routes still need in FABRIC_GW_SSRF_ALLOW_PRIVATE_UPSTREAM_HOSTS, a comma-separated list of exact hostnames or *. wildcards. The allow list has no effect unless blocking is on. The same rule applies to the gateway’s calls to the embedder. If you turn on blocking, also add the embedder’s host name to FABRIC_GW_SSRF_ALLOW_PRIVATE_UPSTREAM_HOSTS, exactly as it appears in your tool search embedder URL. The chart’s embedder Service is fabric-gateway-embedder. Otherwise, MCP tool search fails.
  • The gateway’s own authentication calls. OAuth and OIDC discovery, client registration, token exchange, and JWT Auth key fetches don’t reach private addresses unless you list the host in FABRIC_GW_SSRF_ALLOW_PRIVATE_HOSTS. This list is empty by default and doesn’t affect upstreams. Add a host only if, for example, your identity provider is on your private network.

Set these variables under env.vars. The chart’s values file sets only FABRIC_GW_SSRF_ALLOW_PRIVATE_HOSTS, so add the others yourself. See Helm values.

Ports inside the cluster

These ports are used only between pods in the cluster. You don’t need egress rules for them, but a NetworkPolicy must allow them.

FromToPortPurpose
Clients in the cluster, or your ingressGateway Service8080 (plain HTTP)Request traffic
KubeletGateway container8080, path /healthzLiveness, readiness, and startup probes
Gateway podsEmbedder Service80, forwarded to 8080Tool-search embeddings
Gateway containerCollector sidecar, same pod4317 (gRPC), 4318 (HTTP)OTLP from the gateway
Collector sidecarItself and the embedder8888, and the embedder’s /metricsMetrics scrape
KubeletCollector sidecar13133Collector health checks
Gateway podsOther gateway pods3320 (data), 3322 (gossip)Distributed secret store only
Gateway containerOPA sidecar, same pod8181, on loopback onlyPolicy decisions, only with opa.enabled
KubeletOPA sidecar8181, path /healthReadiness check, only with opa.enabled
Callers of the egress proxyGateway container8090 (llmEgress.port)Transponder LLM egress callback, only with llmEgress.enabled

With opa.enabled, gateway pods don’t become ready. OPA listens only on loopback, so the kubelet’s readiness check on port 8181 fails. See Helm values.

With the distributed store, the chart adds a NetworkPolicy that limits ports 3320 and 3322 to gateway pods. That policy allows no other traffic into gateway pods, so on a cluster that enforces NetworkPolicy it also blocks client traffic to port 8080. Add a rule for 8080 to your values file:

distributed:
networkPolicy:
extraIngress:
- ports:
- protocol: TCP
port: 8080

This rule allows traffic to port 8080 from any source. To limit it to your ingress controller, add a from selector to the rule.

With embedder.networkPolicy.enabled, only gateway pods can reach the embedder. The policy limits ingress only, to the embedder’s container port 8080, and doesn’t restrict egress. In embedder.networkPolicy.extraIngress, use port 8080, not the Service port 80.

For how clients outside the cluster reach port 8080, see Expose the gateway.

Next steps