Open-source LLM rails framework with Colang DSL + LLM-based safety railsNeMo Guardrails is NVIDIA's open-source toolkit for adding programmable rails to LLM applications. It's strongest as a policy framework (Colang DSL, dialogue-flow integration) rather than a fast inline firewall. Pattern-only heuristics target adversarial-suffix attacks (GCG-style); natural-language injection coverage requires the LLM-based self-check rails. We published a public head-to-head measurement on a 100-prompt corpus.
Competitor data (GitHub stars, downloads, feature counts, funding / acquisition status) verified as of 2026-04-28. EvalGuard's own counts are sourced live from the drift-checked registry.
Coverage at a glance
Where both platforms publish a number, here's the gap. Our values come straight from the drift-checked registry; NeMo Guardrails (NVIDIA)'s are quoted as published.
| Feature | EvalGuard | NeMo Guardrails (NVIDIA) |
|---|---|---|
| Detection Recall (pattern-only, same 100-prompt corpus) | 100% (reproducible) | 0% (different attack class by design) |
| p95 Latency (pattern-only) | 3.58 ms | 3280 ms (CPU GPT-2 inference) |
| Cost per scan | $0 (regex + cosine sim) | $0 pattern-only / $/req with LLM rails |
| Public head-to-head | Yes (`benchmarks/head-to-head-nemo-guardrails.md`) | First independent comparison |
| Programmable Policy DSL | Visual rule builder + JSON | Colang DSL (more expressive) |
| Dependencies | Pure TypeScript, zero | torch + transformers + GPT-2 weights (~500 MB) |
| Eval Platform | 188 scorers + 33 frameworks | No |
| Red Team Plugins | 249 | No |
| Agent Tracing | OpenTelemetry | No |
| Cost Analytics | Yes | No |
| Compliance Frameworks | 33 | No |
| Open Source | Apache 2.0 (4 packages on npm) | Apache 2.0 |
Start free. No credit card required. Migrate in minutes.