Architecture in five claims.
Hub-and-spoke through one controller
All overlay traffic between your machines passes through a controller you choose (ours, or one you self-host). The controller is the single place to apply policy. Peers don't need to discover each other.
Per-tenant Linux network namespace
Each organization runs in its own kernel network namespace on the controller. Two tenants on the same controller cannot see each other's overlay traffic, even with identical overlay IP ranges.
iptables-enforced ACLs
Allow/deny policies are written into the FORWARD chain of the tenant's namespace. Rules apply at the kernel level — when a policy says deny, the kernel drops the packet before it reaches the destination.
Encrypted hub keys at rest
Each tenant's hub private key is generated server-side and stored AES-256-GCM encrypted in PostgreSQL. The encryption key is read from an environment variable; the controller refuses to start if it's unset.
No internet egress through the hub
By design and verified: the tenant namespace has no MASQUERADE rules, no SNAT, and no path to public networks. The hub forwards overlay traffic between your peers only. Peers reach the internet via their own local network.