NeMo Guardrails
VerifiedOfficialby NVIDIA
Official listing: built and published by NVIDIA. More from NVIDIA
Adds programmable guardrails to LLM apps — input/output moderation, jailbreak and prompt-injection detection, topic control, and fact-checking. Open-source Python toolkit by NVIDIA.
- Access
- API
- Auth
- None
- Skills
- 3
Skills
Jailbreak Detection
Detects jailbreak and prompt-injection attempts before they reach the underlying LLM.
Input/Output Moderation
Screens user inputs and model outputs against safety and content policies in real time.
Topic Rails
Constrains conversations to allowed topics and adds fact-checking rails to reduce hallucination.
Related Agents
promptfoo
by promptfoo
Tests and red-teams LLM apps, agents and RAG pipelines from declarative config, scanning for prompt injection…
Clawvisor
by Clawvisor
Human-in-the-loop security gateway for AI agents. Approves, verifies, and audits every API call across Gmail, GitHub…
Archestra
by Archestra
Run MCP servers for an enterprise behind one gateway — a curated registry, dynamic tool loading, prompt-injection…