The connection model
A worker you run lives in your environment and is outbound-only: it dials the control plane over TLS, and nothing dials into it. There is no inbound listener to expose, no port to open toward your cluster, and no ingress rule to write for the platform’s benefit. That has a practical consequence worth stating: an agent can work inside a private network that accepts no inbound traffic at all, because the connection is always established from your side.Inbound events — something in your estate telling an agent that a thing happened — arrive at an
endpoint on the platform side that your system posts to, under Integrations → Endpoints. Your
network still only makes outbound connections.
Control-plane egress requirements
For the control-plane connection specifically, allow outbound HTTPS to:
Prefer a hostname-based egress rule over pinning resolved IP addresses — the endpoint sits behind
cloud load balancing, and its addresses can change.
This is the control-plane connection only. Integrations, MCP servers, model providers, and anything
else a self-hosted agent is configured to call are separate outbound destinations, decided by that
agent’s own configuration rather than by this page.
Proxy configuration
If your environment routes outbound traffic through an explicit proxy, setHTTPS_PROXY through
the chart’s env value (see
Deploy a self-hosted agent). The worker’s
HTTP client reads the standard proxy environment variables. Set NO_PROXY as well for any
destination that should bypass the proxy — typically in-cluster ones such as .svc,
.cluster.local and the Kubernetes API, so in-cluster MCP servers and cluster access skip it.
The worker receives its work over a long-lived streaming connection (server-sent events), so the
proxy must not buffer responses or cut idle connections short.
TLS inspection
If that proxy terminates and re-signs TLS, the worker needs to trust the proxy’s CA on top of the public roots. PointcaBundle.existingConfigMap at a ConfigMap holding it — any key name works,
and every file the ConfigMap holds is trusted, so one or more CA certificates can go in together:
caBundle.existingConfigMap: internal-ca. The worker adds this certificate to its public
trust roots; it does not replace them. Pods read the CA at start-up, so restart them after changing
the ConfigMap. See A run fails with a certificate
error for the symptoms.
Verifying it
The clearest signal is the agent itself: if it never shows online in Fleet, the connection is not completing — see Proving it works. To check reachability from inside your network before installing anything, confirm the connection opens from a pod in the target namespace:--env if you don’t use a proxy. Any HTTP status means you reached the control plane; a
timeout or refusal points at the firewall or proxy rule; a certificate error means a TLS-inspecting
proxy is in the path (see TLS inspection).
What an agent can reach on the way out
An agent’s reach is decided by its tool surface as well as by its network, and the tool surface is the decisive one: an agent’s capabilities are governed centrally, so a firewall rule alone does not narrow what it may do. The two work together — platform controls govern the supported tool paths, and network policy constrains the destinations the worker can reach at all. Two paths are governed — a built-in integration, or the MCP Gateway — and these are the controls over them:What crosses the wire
Outbound from the worker: the run ledger and the evidence an agent produces — its inputs, tool calls, reasoning, outputs, and cost — with run-scoped secrets masked before they leave. See Data handling & redaction for what that covers and what it does not. Inbound to the worker: work to do, and the credentials for the run doing it. Nothing else, and nothing unsolicited.Next steps
MCP Gateway
The governed path for your own tools.
Data handling & redaction
What is in the traffic that does leave.
Agent isolation & tenancy
The boundary the connection lands inside.
Standards & Guardrails
Refusing a call before it leaves.