OpenAI Guardrails (Python)

OpenAI Guardrails for Python is an open-source library from OpenAI, currently in preview, that wraps the OpenAI Python client to validate inputs and outputs of LLM calls. Pipelines are configured in a web wizard, exported as a JSON file, and loaded by the client at runtime. It also integrates with the OpenAI Agents SDK.

What It Is

The library is a drop-in replacement layer for the OpenAI client. Developers use GuardrailsOpenAI in place of OpenAI, and requests through Chat Completions or Responses APIs are checked against the configured guardrails. When a check trips, a GuardrailTripwireTriggered exception is raised so the application can block the message. Async and Azure clients are covered in the docs.

Available Checks

Guardrails are grouped into input, output and agentic stages. Documented checks include keyword filtering, competitor mentions, moderation, URL allow/block lists, secret key detection, PII detection and masking, hallucination detection against vector-store reference material, jailbreak detection, prompt injection detection for misaligned tool calls, NSFW text, off-topic prompts, and custom prompt checks. The configuration site lists estimated latency for each check and the type of mechanism it uses, such as regex, moderation API, LLM, or hybrid.

Agents SDK and Evaluations

GuardrailAgent applies the same exported configuration to agents built with the OpenAI Agents SDK, raising input or output tripwire exceptions. An evaluation module measures guardrail performance on labeled JSONL datasets using the same config, with optional benchmark dependencies for model comparisons.

Setup Path

Install with pip install openai-guardrails on Python 3.11+, export a configuration from the wizard, and set OPENAI_API_KEY. The Contains PII check requires a spaCy model to be installed. Guardrails call paid OpenAI APIs, and developers are responsible for those charges and for handling sensitive content appropriately. The code is MIT licensed.