AI ACTION CONTROL · LAYER 4

The Action Capsule.

A Kubernetes-native container plus in-container sandbox, around one model, agent or MCP server. Paired Gateway as the only egress.

A chatbot answers. An agent acts. When the runtime starts calling tools, writing to systems of record or running long tasks, you need a Capsule around it.

A standard Kubernetes pod with default-deny network policy contains a Kosmoy Action Capsule — the Docker container plus the in-container sandbox built on Linux primitives. The Capsule’s only egress is the paired Kosmoy AI Gateway, which fronts approved models, MCP servers, A2A peers and HTTPS APIs.KUBERNETES PODstandard pod · Helm / ArgoCDL3 / L4 · default-deny egress · pinned netnsKOSMOY ACTION CAPSULEDocker container + in-container sandboxLinux namespacescgroups v2Seccomp · capabilitiesLandlockAppArmor / SELinuxprofileL7 · in-pod forward proxyHTTP allow-list · TLS pinning · method & path · DNS pinned · auditRUNTIMEModel · Agent task · MCP serverCPU · memory · GPU · timeout · concurrencyrun-scoped lease · JIT credentialsNo provider keys reach this layer.only egressKOSMOY AI GATEWAY · PAIREDpolicy point · auth · routing · loggingApproved LLMsOpenAI · Anthropic · private · SLMApproved MCP serverspublic · private · scoped toolsApproved A2A peersagent-to-agent · run-scopedApproved HTTPS APIshost + path allow-list
The Capsule is the boundary. Default-deny egress at the network layer, an L7 proxy at the application layer, the paired AI Gateway as the only way out.

Five pillars.

Kubernetes-native

Helm-installable. ArgoCD-friendly. EKS, AKS, GKE, OpenShift, on-prem. No host or node changes.

Container + sandbox in one

A Capsule combines a Docker container with an in-container sandbox built on Linux primitives — namespaces, cgroups v2, Seccomp, Linux capabilities, Landlock, AppArmor / SELinux.

L3 / L4 + L7 controls

Default-deny egress at the network layer. An in-pod L7 proxy enforces the HTTP allow-list, TLS pinning and DNS pinning at the application layer.

Pre-flight authorisation

Every run is admitted before it executes. Run-scoped lease. JIT credentials issued for the run, revoked at the end.

Live kill switch

Stop a misbehaving Capsule mid-run. Cancel future schedules. Every action captured as evidence.


Capsule + Gateway pairing.

Each Capsule is paired with one Gateway. The Gateway defines what the Capsule can reach. A Gateway can also serve apps without a Capsule, so Gateways scale wider than Capsules.

Two patterns. First, map existing applications and route them through a Kosmoy AI Gateway. Second, run the runtime inside an Action Capsule and make the paired Gateway its only egress.Phase 1 · Gateway aloneMapdiscover existing AI callsCustom appRAG · chatbot · agent workflowKosmoy AI Gatewaypolicy · budget · logsApproved external AILLMs · MCP · A2A · APIsPhase 2 · Capsule + GatewayMission Controlsupervise · approve · stopAction Capsuleagent · model · MCP runtimePaired Gatewaythe only egressApproved egressmodels · MCP · A2A · HTTPS
Map first. Govern through the Gateway. Add Capsules when the runtime itself must be contained.

Module questions, answered straight.

What is an Action Capsule?

A Kosmoy Action Capsule is a Docker container plus an in-container sandbox, deployed inside a standard Kubernetes pod. The Capsule runs one model, one agent task or one MCP server. The only egress is the paired Kosmoy AI Gateway.

How is the Capsule isolated?

Application-level isolation built on standard Linux primitives — namespaces, cgroups v2, Seccomp-BPF, Linux capabilities, Landlock and AppArmor / SELinux profiles. No node patches, no custom container runtime.

How is network egress controlled?

Two layers. At L3 / L4 a Kubernetes NetworkPolicy applies default-deny egress and the netns is pinned. At L7 an in-pod forward proxy enforces an HTTP host + path allow-list, TLS pinning, method rules and pinned DNS. The only egress is the paired Kosmoy AI Gateway.

Why pair every Capsule with a Gateway?

Containment without controlled egress is just a wall with no door. The Gateway is the door — it logs, authorizes and policy-checks every step out.

How are compute resources enforced?

CPU, memory, GPU, timeout and concurrency limits are defined at Capsule creation. cgroups v2 enforce CPU / memory; the Action Plane enforces timeout and concurrency.

How do you contain an AI agent that misbehaves?

Containment has to hold even when the agent's own logic is compromised, so it is enforced at the runtime boundary, not in the prompt. Inside a Capsule the agent keeps its productivity but cannot reach systems outside its authorized scope; the paired Gateway is the only egress; credentials are just-in-time and per task; and the kill switch stops the process itself — no cooperation from the agent required.

Enterprise proof

Built for environments where AI cannot become another unmanaged SaaS silo.

Kosmoy works with regulated and operationally critical organizations, including Banca d’Italia and Leonardo.

Banca d'Italia
Leonardo
  • Runs in your Kubernetes
  • Azure · AWS · GCP · on-prem
  • LLM · MCP · A2A
  • No mandatory vendor-hosted control plane

See an Action Capsule running.

Walk through Capsule, paired Gateway and the live kill switch.

Or email sales@kosmoy.com.