# Omnia vs NeMo Guardrails

> Omnia applies rules to tool-call argument values and caller identity on every model provider it runs. NeMo Guardrails checks tool calls on some providers only, and is a library you wire into each agent.

Category: Gateways and guardrails. Checked against NVIDIA’s published material on 2026-10-22. Canonical page: https://www.altairalabs.ai/vs/nemo-guardrails/

## NeMo Guardrails checks tool calls on some providers, where Omnia checks the argument values on every one it runs.

Omnia’s checks live in the runtime, so they apply to every agent and every provider it runs. NeMo Guardrails is an open toolkit of dialogue and tool rails that you wire into each agent.

## When it matters: A Claude agent on Bedrock is about to issue a refund.

**On Omnia:** The refund call passes through the policy broker like every other tool call. The amount is over the caller’s limit, so the call is refused before it runs. (Omnia docs: https://omnia.altairalabs.ai/how-to/security/configure-tool-policies/)

**On NeMo Guardrails:** Tool-calling rails don’t yet support the OpenAI Responses API, Anthropic, Gemini or Bedrock. (source: https://docs.nvidia.com/nemo/guardrails/configure-guardrails/guardrail-catalog/tool-calling)

## Values, not just shape

A CEL rule reads the arguments’ values and the caller’s claims: refunds under £500, to the caller’s own account, by agents allowed to refund. *(enterprise)* (Omnia docs: https://omnia.altairalabs.ai/how-to/security/configure-tool-policies/)

NeMo Guardrails: Arguments must validate against the tool’s declared JSON Schema; rules on values aren’t documented. (source: https://docs.nvidia.com/nemo/guardrails/configure-guardrails/guardrail-catalog/tool-calling)

## Every provider Omnia runs

Policy sits at the tool call, so the same rule applies whichever model asked: Anthropic, OpenAI, Gemini, Bedrock, Vertex, Ollama or vLLM. *(enterprise)* (Omnia docs: https://omnia.altairalabs.ai/how-to/security/configure-tool-policies/)

NeMo Guardrails: Tool rails don’t yet cover Anthropic, Gemini or Bedrock. (source: https://docs.nvidia.com/nemo/guardrails/configure-guardrails/guardrail-catalog/tool-calling)

## Built into the platform

Policy, identity, redaction, testing and rollout come with the runtime. The agent’s code stays as it is. (Omnia docs: https://omnia.altairalabs.ai/explanation/security/policy-engine/)

NeMo Guardrails: A Python library or self-hosted server; the production microservice needs an NVIDIA AI Enterprise licence. (source: https://docs.nvidia.com/nemo/guardrails/about-nemo-guardrails-library/overview)

## At a glance

| Topic | Omnia | Omnia docs | NeMo Guardrails | NeMo Guardrails docs | Source |
|---|---|---|---|---|---|
| Open source | PromptKit and PromptArena, Apache 2.0. | https://promptarena.altairalabs.ai/arena/explanation/scenario-design/ | Apache 2.0. | documented | https://github.com/NVIDIA-NeMo/Guardrails |
| Fails closed on tool calls | Yes. *(enterprise)* | https://omnia.altairalabs.ai/how-to/security/configure-tool-policies/ | Yes, for supported APIs. | partly documented | https://docs.nvidia.com/nemo/guardrails/configure-guardrails/guardrail-catalog/tool-calling |
| Rules on argument values | Yes. *(enterprise)* | https://omnia.altairalabs.ai/how-to/security/configure-tool-policies/ | Schema validation. | partly documented | https://docs.nvidia.com/nemo/guardrails/configure-guardrails/guardrail-catalog/tool-calling |
| End-user identity in policy | Yes. *(enterprise)* | https://omnia.altairalabs.ai/reference/policies/toolpolicy/ | Not documented. | not documented | https://docs.nvidia.com/nemo/guardrails/about-nemo-guardrails-library/overview |
| Red teaming before release | Adversarial callers gate the release. | https://promptarena.altairalabs.ai/arena/how-to/voice/voice-red-team/ | A separate tool, garak. | partly documented | https://github.com/NVIDIA/garak |

## Questions

**Can I keep NeMo Guardrails?**

Its dialogue rails can stay in the agent’s own code, and Omnia’s policy still applies at each tool call.