Omnia · the self-hostable AgentOS

The platform you own, running in the cluster you already trust.

Run AI agents in production while your systems, your data, and your audit trail stay in-house, never routed through a vendor's cloud. Omnia is Kubernetes-native and model-agnostic, and it is built out of things you already operate — containers, OCI, OpenTelemetry, Postgres. There is nothing in here your platform team has to learn from scratch.

Talk to a founder Start free, self-hosted →

See it end to end

From a plain-language brief to a governed agent in production.

Five steps, one artifact. The pack that passes the arena is the same pack that gets installed into the cluster and answers real requests.

claude · code

recording

Recording drops in here — save define.cast to /public/atlas/casts/

Describe the agent in plain language. One claude -p call scaffolds a full PromptArena kit from the authoring skill PromptArena ships.

How it fits together

Nothing in here you don’t already operate.

Every layer maps to something on your cluster today, which is why this clears a platform review instead of starting an argument. It all runs inside your boundary, and the model provider is a config line you can change.

Own · in your cluster · your IP

the spine User ↔ Agent ↔ Your proprietary logic

Define

PromptPack

Open, declarative spec for an agent: defined, versioned, portable.

Prove

PromptArena

Self-play against the agent's real tools, in-cluster, wired into the deploy gate.

Run

PromptKit

Lean Go runtime for text, voice, and video. Where the agent actually runs.

Operate

Omnia THE PLATFORM

Memory · telemetry · governance · compliance. The services every agent shares. Free + Enterprise.

Substrate · already in your stack

KubernetesOCIMCPOpenTelemetryPostgres

Reach · your systems of record

Capabilities — the agent’s hands

MCP tools, packaged as Kubernetes microservices, reach into the systems you already run.

Only the tools you grant
Pre-defined operations, not raw access
Every call audited

The model · pluggable

Works with any provider

One config across all of them. Swap or run several at once, no re-platforming.

OpenAIAnthropicGeminiOllama / vLLM …or grow your own

Your agents are your IP: the logic, the memory, the policy, all running in your cluster. The model is the one piece you never have to commit to: bring any provider, swap whenever, run several side by side.

The control plane

Operate agents the way you operate everything else.

A real control plane: the whole fleet with drift between desired and running state, the live path a request takes through the system, and each agent’s health down to the spec it’s running.

omnia.your-cluster.internal

design

Fleet overview — every agent, its running replicas, and the drift between desired and live state, at a glance.

Interface design. The CLI recordings above are real runs.

Yours end to end

Nothing you have to take on trust.

Your SRE can inspect it, run it in their own cluster, and clear the lock-in veto before anyone signs. That's what makes the open layers a trust mechanism, not a black box.

Runs in your tenancy

Your credentials, your systems of record, your VPC. You decide what context ever reaches the model.

Governed by default

RBAC, approval gates, and append-only audit exported to your SIEM — the artifacts your security reviewer asks for.

Model-agnostic

OpenAI, Anthropic, Gemini, Azure, Ollama, vLLM, or a model you grow yourself. Swap anytime; the agents are the part you keep.

Start with the part that's free

Two ways in. Both free.

Both run in your own cluster, free to start. Don't book a call — install one and see.

Run it · Omnia Free

want to deploy & govern

Stand the platform up in your own cluster and run a governed agent against a sandbox system. Reach for Enterprise when it touches production.

$ helm install omnia altairalabs/omnia

Prove it · PromptArena

want to test & break things

Run adversarial and voice personas against your agent in CI. No API keys, any provider. The fastest way to see it hold or crack.

$ npm i -g @altairalabs/promptarena

See PromptArena →