Guardrails

Guardrails is a Python framework for making LLM applications more reliable in two ways: it runs input and output guards that detect, quantify, and mitigate specific categories of risk, and it generates structured data from an LLM.

Guards are built from validators — individual risk measures published on Guardrails Hub and installed as Python packages. You compose one or more validators into a Guard and give each an on-fail action, such as raising an exception, then call guard.validate() on text passing into or out of the model. Examples include regular-expression matching, competitor checks, and toxic-language detection.

For structured output, you define the shape you want as a Pydantic model and create a guard from it. The guard then calls the LLM so the response matches that class, either through function calling on models that support it or by adding the expected output schema to the prompt for models that do not.

Guardrails also runs as a standalone service. guardrails start serves configured guards over a REST API, which you can call with the Guardrails client or by pointing an OpenAI SDK at the guard's own endpoint. It can be used with proprietary and open-source LLMs, from Python and JavaScript, and you can write and contribute your own validators.

Validators are moving to standard PyPI packages installed directly with pip, and hosted remote inferencing is being discontinued, with a planned cutoff of August 25, 2026.