Restricted networks
Everything on this page is optional and off by default. A cluster with ordinary outbound internet access needs none of it.
What the agent actually needs
Section titled “What the agent actually needs”| Value | |
|---|---|
| Direction | Outbound only. The agent dials out; nothing ever connects to it. There is no Service, no Ingress and no container port |
| Destination | Seylo’s agent host — the one in the server URL. On Seylo’s hosted service, agent.seylo.in for agents installed with today’s commands; portal.seylo.in for agents installed earlier |
| Port | 443 for wss:// |
| Protocol | HTTPS upgraded to a WebSocket (GET /agent/connect) |
| Connections | One, long-lived, re-established with backoff when it drops. From agent 0.2.17 the agent pings it every 15 seconds, so a drop is noticed in about 25 seconds. If a proxy on the path drops WebSocket pings, set connection.pingSeconds=0 |
| Also needed | In-cluster DNS for Seylo’s hostname, and the Kubernetes API server via the ServiceAccount |
| Not needed | Any inbound rule, any public API endpoint, any cloud credential, any registry access at runtime — only at image pull |
A single egress allow-rule to agent.seylo.in:443 is sufficient.
-
Agents installed earlier dial
portal.seylo.in:443, and keep working there. Keep that rule for as long as any of your agents still dials it, or move them and then remove it. To see which host an agent dials:Terminal window kubectl -n seylo-agent get deploy seylo-agent -o jsonpath='{.spec.template.spec.containers[0].env[?(@.name=="SERVER_URL")].value}'
If you use NetworkPolicies, the agent’s namespace needs an egress policy permitting that host plus DNS to CoreDNS on 53.
The agent host serves agents only
Section titled “The agent host serves agents only”agent.seylo.in answers one path, /agent/connect, and only to a WebSocket
upgrade from a Seylo agent. Opened in a browser or with a plain curl, it
answers HTTP 426 with this, which is expected:
/agent/connect accepts only a WebSocket upgrade from a Seylo agent. Install the agent from Admin > Clusters; it dials this address itself.Any other path on that host answers 404. The portal, sign-in and the API stay
on portal.seylo.in.
Nothing in the security model changes with the host. The connection is
TLS end to end, on a certificate for agent.seylo.in that the agent verifies,
and the agent still only dials out, authenticates with its own per-cluster
credential, and runs only the commands it allows. Your cluster still accepts
nothing inbound.
Through an HTTP forward proxy
Section titled “Through an HTTP forward proxy”The agent honours the conventional variables and tunnels its WebSocket through
the proxy with CONNECT. The proxy’s allow-list must permit CONNECT to
agent.seylo.in:443 (and to portal.seylo.in:443 for an agent installed
earlier).
| Variable | Helm value | Meaning |
|---|---|---|
AGENT_PROXY_URL |
proxy.url or proxy.existingSecret |
Explicit override. Wins over the ambient variables, so the chart can set it without fighting a cluster-wide HTTPS_PROXY injected by a webhook |
HTTPS_PROXY / https_proxy |
— | Standard. Uppercase wins over lowercase |
HTTP_PROXY / http_proxy |
— | Standard, used when neither of the above is set |
NO_PROXY / no_proxy |
proxy.noProxy |
Bypass list. A match wins over everything, including the explicit override |
Precedence, exactly:
NO_PROXYmatches the server host → direct connection. This beats every proxy variable, because “do not proxy this host” is the more specific instruction.AGENT_PROXY_URLHTTPS_PROXY, thenhttps_proxyHTTP_PROXY, thenhttp_proxy- Nothing set → direct connection.
A proxy value that cannot be parsed as an http(s) URL is not silently
ignored. It is reported as config-proxy-url-malformed, and the agent says it
is connecting directly instead.
NO_PROXY grammar
Section titled “NO_PROXY grammar”Comma- or whitespace-separated, case-insensitive.
| Entry | Matches |
|---|---|
* |
Everything — never use the proxy |
tool.corp.local |
That exact host and any subdomain of it |
.corp.local |
corp.local and every subdomain |
corp.local |
corp.local and tool.corp.local, but not notcorp.local — the label boundary is respected |
tool.corp.local:8443 |
That host only on that port |
10.1.2.3, [fd00::1] |
IP literals, exact |
10.0.0.0/8, fd00::/8 |
CIDR — only when the target host is itself an IP literal |
| empty / whitespace | Nothing |
Proxy authentication
Section titled “Proxy authentication”Taken from the userinfo in the proxy URL and sent as Proxy-Authorization:
--set proxy.url='http://svc-agent:s3cr3t@proxy.corp.local:3128'Prefer an existing Secret. A Deployment’s environment is readable by anyone who can read the Deployment, which is a wider set of people than you think:
--set proxy.existingSecret=agent-proxy-urlTLS inspection and private certificate authorities
Section titled “TLS inspection and private certificate authorities”If your egress proxy terminates and re-signs TLS, the agent sees a certificate
signed by your CA, not a public one. That produces tls-unknown-ca.
The fix is to give the agent your CA bundle:
# Let the chart create the Secret from a local file.--set-file ca.pem=/path/to/corp-ca.pem
# Or reference one that already exists, e.g. managed by an external secret store.--set ca.existingSecret=corp-ca-bundleAir-gapped clusters
Section titled “Air-gapped clusters”A fully air-gapped cluster — no egress to Seylo at all — cannot be managed by the agent. The agent’s whole design is one outbound connection; with no path out, there is nothing to invert.
Mirroring the image into an internal registry solves the image pull, not the connection. If your cluster can reach Seylo through a proxy, you are not air-gapped in the sense that matters here, and the proxy configuration above is what you need.
Mirroring the image
Section titled “Mirroring the image”If policy requires images to come from your own registry, the image and chart can both be mirrored. The digest is what makes this safe to verify: copy the image, record the digest, and confirm it is unchanged on the far side before installing.
Pass your mirror to the chart with --set image.repository=…, plus an image
pull secret if your registry requires authentication. The published upstream
image is public specifically so that a mirror is a choice rather than a
requirement.