Send pool traffic through your proxy
Many organizations require everything leaving a cluster to pass through a corporate proxy. A private pool can do that: point it at your proxy once, and the Operator, Armor and the SIEM collector use it. You choose which kinds of traffic go through it, and which MCP servers do.
A pool with no proxy configured behaves exactly as it did before — everything connects straight from the cluster.
Quick Start: What do you want to do?
| I want to... | Jump to |
|---|---|
| Send my pool's traffic through a proxy | Set up the proxy |
| Choose what goes through it and what doesn't | Choose which traffic uses the proxy |
| Keep some hosts off the proxy | Always direct |
| Send one MCP server through the proxy | Per-server network path |
| Stop a bad proxy cutting off my pool | Fail open |
Set up the proxy
-
Open your pool from Deployments and go to the Configuration tab.
-
Find the Outbound Network card — How the operator and Armor reach services outside the cluster — and select Edit outbound network.
-
Under Outbound traffic, choose Through a proxy.
-
Choose a Proxy type:
Type What it does HTTP Tunnels HTTPS traffic with CONNECT. HTTPS Like HTTP, with TLS to the proxy itself. SOCKS5 A TCP tunnel; the proxy resolves names. -
Enter the Proxy address as host and port — for example
proxy.example.com:3128. Don't includehttp://orsocks5://. -
If your proxy requires sign-in, fill in Username and Password. Both are optional; a password needs a username. Once saved, the field shows Password saved and you can leave it empty to keep it.
-
Save the pool configuration.
With all three proxy types the proxy resolves the destination's name, so the cluster itself needs no external DNS.
A proxy change reaches Armor as new environment settings, so its pods restart to pick them up. Expect a brief interruption, and make the change in a quiet window.
Choose which traffic uses the proxy
Three kinds of traffic are controlled separately, so you can send the pool's own connections through the proxy while leaving others direct.
| Setting | Covers |
|---|---|
| Control plane connection | The Operator's link to Cequence: configuration, heartbeat and logs. |
| Sign-in and token checks | JWKS, OIDC discovery, OAuth token and client-registration calls, and client metadata documents. |
| SIEM export | Audit and activity export from the SIEM collector to Splunk, Datadog or Grafana. |
Tool calls to MCP servers are not in this list. They are decided per server — see Per-server network path. By default an MCP server connects directly even when the pool has a proxy.
Only HTTP destinations — Splunk HEC, Datadog, Grafana and OTLP over HTTP — can go through the proxy. Syslog and Splunk forwarder destinations always connect directly.
Sign-in traffic follows its own toggle regardless of how a server is set. If Sign-in and token checks is on, a server that connects directly still has its OAuth and token calls go through the proxy.
Always direct
Some destinations shouldn't go through the proxy at all. List them under Always direct as hosts, domains (.corp.internal) or CIDR ranges, separated by commas — up to 50 entries, and no spaces inside an entry.
Services inside the cluster are always direct and never need listing.
Allowed outbound ports
Allowed outbound ports applies when NetworkPolicies are turned on for the pool. The defaults are 80 and 443, and you can list up to 20 ports. When the pool has a proxy, its port is added automatically — the field notes this rather than making you add it.
Fail open
Fail open decides what happens when the Operator can't reach Cequence through the proxy. Leave it on and the Operator connects directly instead and reports that it did, so a misconfigured proxy can't cut your pool off from management. Turn it off only when policy requires that all traffic go through the proxy, even at the cost of losing contact.
A new proxy is tested before the Operator commits to it, so a setting that doesn't work is caught rather than applied.
Per-server network path
Each MCP server in the pool chooses its own path. This matters because a corporate proxy often can't reach internal services, while an external SaaS API may only be reachable through it.
| Choice | What happens |
|---|---|
| Through the pool's proxy | Tool calls go through the proxy. |
| Connect directly | Tool calls go straight from the cluster. For internal servers the proxy can't reach. |
Set it under Network path when you add an MCP server to the pool, or afterwards from the server's configuration in the pool. A server using the proxy shows a Via proxy badge.
If the pool has no proxy, the selector says so — This pool has no proxy, so tool calls connect directly. If you set one, this server will use it — and the badge reads Proxy (pool has none).
If several servers share a hostname and any one of them uses the proxy, all traffic to that host goes through the proxy. A choice to use the proxy is never bypassed.
When you change or remove the proxy
The pool tells you which servers are affected before you save. Removing a proxy that servers use warns that tool calls from those servers go straight from the cluster, and asks you to check those servers accept direct connections and that their ports are in Allowed outbound ports. Pointing the pool at a different proxy warns that those servers' calls will go through the new one.
How this relates to the bootstrap proxy secret
If you install a pool in a cluster with no direct internet access, you give the Operator a proxy before it can reach Cequence at all — that's the aigw-proxy-config secret in Air-gapped clusters.
The two work together:
- The secret is how the Operator gets out at startup, before it can fetch any pool configuration. It is read once, at startup, so changing it means restarting the Operator.
- The Outbound Network card takes over once the Operator is talking to Cequence, and covers Armor, the SIEM collector and MCP servers as well.
A pool with no Outbound Network setting keeps using the Operator's bootstrap proxy. Set the card and it takes precedence.
TLS inspection is handled only by the secret: if your proxy re-signs certificates, its CA certificate goes in the secret as caCert.
Tips
- Start with the pool's own traffic. Turn on Control plane connection first, confirm the pool stays healthy, then move MCP servers over one at a time.
- Leave Fail open on unless policy forbids it. It's the difference between a bad proxy causing a problem and a bad proxy causing an outage.
- Put internal services in Always direct rather than relying on each server's network path — it covers the Operator and SIEM collector too.
- Change the proxy in a quiet window, since Armor restarts to pick it up.
- Check the ports. When NetworkPolicies are on, a server reachable on a non-standard port needs that port in Allowed outbound ports.
Cequence AI Gateway