Self-Hosted Kubernetes

Run AIControls entirely inside your own cluster. Every audit event, policy decision, and token budget stays in your infrastructure — no data leaves your environment.

What self-hosted K8s mode is

The self-hosted deployment runs the AIControls binary as a Kubernetes Deployment, backed by a PVC-mounted SQLite database and configured via Helm. There is no dependency on the Nirmata SaaS platform — policy sources, audit storage, and identity validation are all cluster-local.

You get the same governance pipeline as the managed offering: MCP proxy, LLM proxy, CEL policies, HITL approval, token budgets, audit log, and the admin UI. The difference is operational: you own the infrastructure, control the upgrade cadence, and choose where data lives.

Architecture

┌──────────────────────────────────────────────────────────────────┐
│  Kubernetes Cluster                                              │
│                                                                  │
│  ┌─────────────────────┐     admission webhook                   │
│  │  Kyverno            │──────────────────-────┐                 │
│  │  (optional)         │                       │                 │
│  └─────────────────────┘                       ▼                 │
│                                   ┌────────────────────┐         │
│  ┌─────────────┐   MCP calls      │                    │         │
│  │  Agent Pod  │─────────────────▶│   AIControls       │         │
│  │  (Claude    │                  │                    │         │
│  │   Code /    │   LLM calls      │  ┌──────────────┐  │         │
│  │   Cursor /  │─────────────────▶│  │  CEL Policy  │  │         │
│  │   custom)   │                  │  │  Engine      │  │         │
│  └─────────────┘                  │  └──────────────┘  │         │
│                                   │  ┌──────────────┐  │         │
│                                   │  │  Audit Log   │  │         │
│                                   │  │  (SQLite)    │  │         │
│                                   │  └──────────────┘  │         │
│                                   │  ┌──────────────┐  │         │
│                                   │  │  HITL Queue  │  │         │
│                                   │  └──────────────┘  │         │
│                                   └────────┬───────────┘         │
│                                            │                     │
│                      ┌─────────────────────┼──────────────┐      │
│                      ▼                     ▼              ▼      │
│              ┌──────────────┐   ┌─────────────────┐  ┌──────┐    │
│              │  MCP Servers │   │  Upstream LLMs  │  │ HITL │    │
│              │  (kubectl,   │   │  (Anthropic,    │  │ Chan.│    │
│              │   bash,      │   │   OpenAI,       │  │(Slack│    │
│              │   postgres…) │   │   Bedrock,      │  │/mail)│    │
│              └──────────────┘   │   Ollama/vLLM)  │  └──────┘    │
│                                 └─────────────────┘              │
└──────────────────────────────────────────────────────────────────┘

Ports:

Port	Protocol	Purpose
`:8080`	HTTP	MCP proxy (SSE + stdio) + OpenAI-compatible LLM proxy
`:8081`	HTTP	Admin UI + REST API

warning

PostgreSQL backend is required for production HA installations. replicaCount > 1 is only safe with the PostgreSQL backend — with the default SQLite backend, two replicas will race on the same PVC. If you're staying on SQLite, keep replicaCount: 1 behind a load balancer that routes all traffic to a single pod.