NVIDIA NeMo Guardrails is an open-source toolkit for adding programmable guardrails to applications built on large language models. It sits between application code and the model and offers five types of rails: input, dialogue, retrieval, execution (for tool calls) and output. These address risks such as jailbreaks and prompt injection and keep assistants on topic, and dialogue flows are defined in the Colang modelling language. Developers use it as a Python library, from the command line or as an HTTP guardrails server, and a Dockerfile is provided.
The project is released under the Apache 2.0 licence and maintained by NVIDIA, a US company, in the NVIDIA-NeMo GitHub organisation. NVIDIA controls the project; it is not community-governed. Because organisations install and run the toolkit themselves, the hosting location is their own choice, including infrastructure in the EU. The library collects anonymous telemetry, which can be disabled through environment variables or a do_not_track file. Any models plugged in as checks bring their own hosting and jurisdiction.