Server, CLI and rail library

The NeMo Guardrails CLI is the nemoguardrails command installed with the package: it chats with a config folder, serves configs over HTTP, and the package also ships a library of ready-made rails.

Last updated: 30 Sep, 2026 · NeMo Guardrails 0.24.1

This course used NeMo from Python, one config at a time. The same package runs as a server, and NeMo is one of several guardrail frameworks, which the video compares before choosing it.

The clip names four ways to add guardrails: NeMo Guardrails from NVIDIA, Guardrails AI, Llama Firewall from Meta, and AWS Bedrock Guardrails, a cloud-native service. Choose by use case, and by open source or paid.

Syntax:

nemoguardrails chat --config=.        # talk to a config folder in the terminal
nemoguardrails server --config=configs  # serve every folder in configs/ over HTTP

The commands

import subprocess
import sys

out = subprocess.run([sys.executable, "-m", "nemoguardrails", "--help"],
                     capture_output=True, text=True).stdout
for name in ["chat", "server", "actions-server", "eval", "convert"]:
    print(name, name in out)

chat True
server True
actions-server True
eval True
convert True

• chat opens an interactive conversation with a folder, handy for trying rails by hand.
• server starts an HTTP server with an OpenAI-style chat endpoint, so other applications call your rails.
• actions-server runs actions in a separate process; eval evaluates a config; convert moves Colang 1.0 files to 2.x.

The rail library

import os

import nemoguardrails.library as library

folders = sorted(name for name in os.listdir(os.path.dirname(library.__file__))
                 if not name.startswith("_") and "." not in name)
print(len(folders))
print(", ".join(folders))

32
activefence, ai_defense, attention, autoalign, clavata, cleanlab, content_safety, context_bloat_detection, crowdstrike_aidr, f5, factchecking, fiddler, gcp_moderate_text, gliner, guardrails_ai, hallucination, hf_classifier, injection_detection, jailbreak_detection, llama_guard, pangea, patronusai, policyai, polygraf, privateai, prompt_security, regex, self_check, sensitive_data_detection, topic_safety, trend_micro, utils

self_check is where self check input came from. jailbreak_detection, injection_detection, content_safety, llama_guard and sensitive_data_detection cover what this course wrote by hand, with models trained for the job. Each is switched on as self check input was: a flow name in config.yml, plus whatever model or key it needs.

Guardrail frameworks compared

Where each fits

• NeMo when the rules should be readable files, with your own flows and actions.
• A trained classifier, such as Llama Guard from the rail library, when paraphrased attacks matter more than cost.

Related

• Previous: Groq or Gemini
• Next: Every rail in one config
• Reference: Guardrail catalog