Anex

What is Anex

Tool for running GPU workloads as standard Kubernetes Pods on significantly cheaper cloud providers like Vast.AI and RunPod
Anex is a Virtual Kubelet provider. Deploying it into a cluster registers a virtual node — Pods scheduled there have their containers spawned on a rented machine in the configured cloud provider
A Gateway pod terminates a Wireguard tunnel that remote containers connect to, so cluster traffic can reach pods running on the rented machine
Scheduling stays with the default scheduler. No changes to existing infra required

Anex selects the active provider via cloudProvider.active (vastai, runpod, or mock):

Provider	Notes
`vastai`	Full feature set: machine bans, hardware filters incl. DLPerf, GPU mapping. Wireguard over plain UDP.
`runpod`	Wireguard tunneled over WebSocket (wstunnel) since RunPod blocks UDP. No machine bans; uses a hardcoded GPU price dictionary for filtering.
`mock`	In-memory provider used in tests.

One container per pod (enforced by ValidatingAdmissionPolicy)
No init containers; image command must be set explicitly on the container
kubectl logs / kubectl exec are not available in managed clusters without root CA access
User-space networking on Vast.AI: traffic to the cluster goes through an HTTP proxy or a TCP tunnel
Service resources can’t reach virtual pods directly — use VirtualService (CRD) instead
Service accounts are not propagated to remote pods

Full Kubernetes Service API support (LoadBalancer already partially via NodePort VirtualService)
Service account propagation

Quick start guide to run your first GPU workload with Anex.

Helm chart values and provider configuration

Deploying workloads on the Anex virtual node

Anex deployment examples