any-guardrail provides a unified interface for AI safety guardrails, for example, letting you detect toxic content, jailbreak attempts, and other risks in LLM inputs and outputs. Switch between different guardrail providers, both encoder-based (discriminative) and decoder-based (generative) models like Llama Guard and ShieldGemma, without changing your code.

Some guardrails are extremely customizable, which any-guardrail fully exposes. See the complete list of supported providers and customization examples in our docs.

• Unified API: Switch between evergrowing list of guardrail providers
• Production-ready: Built for real-world LLM applications
• Flexible: Use encoder-based (fast) or decoder-based (customizable) models

• Python 3.11 or newer

Install with pip:

pip install any-guardrail

AnyGuardrail provides a seamless interface for interacting with the guardrail models. It allows you to see a list of all the supported guardrails, and to instantiate each supported guardrail. Here is a full example:

from any_guardrail import AnyGuardrail, GuardrailName, GuardrailOutput

# Initialize guardrail
guardrail = AnyGuardrail.create(GuardrailName.DEEPSET)

# Validate input before sending to your LLM
result: GuardrailOutput = guardrail.validate("How do I hack into a system?")

if not result.valid:
    print(f"Blocked: {result.explanation}")
else:
    # Safe to proceed with LLM call
    response = your_llm(user_input)

Every guardrail returns the same GuardrailOutput shape, so you can swap models without changing application code:

result.valid       # bool verdict — True means the content passed
result.score       # risk score in ~[0, 1], higher = more likely violating (when available)
result.categories  # per-category results: CategoryResult(name, description, triggered, score, severity)
result.explanation # human-readable rationale (judge reasoning, raw generation)
result.action      # provider-recommended action (e.g. "block"), advisory; None if none
result.usage       # provenance: model_id, latency_ms, token counts
result.extra       # guardrail-specific structured extras; result.raw holds the backend payload

flagged = [c.name for c in result.categories if c.triggered]

A machine-readable JSON Schema for this output is published in the repo (generated from the Pydantic models). Reference it at the stable raw URL, pinning a release tag for a specific version:

https://raw.githubusercontent.com/mozilla-ai/any-guardrail/main/schemas/guardrail_output.schema.json

Generative and judge guardrails run against a prompt template and often need a policy, rubric, or criteria. any-guardrail lets you discover and override the default prompt, and fetch author-published policies/rubrics/criteria — all without loading a model:

from any_guardrail import AnyGuardrail, GuardrailName

# Inspect a guardrail's default prompt template
AnyGuardrail.get_prompt(GuardrailName.SELENE).segments["user"]

# Fetch a ready-made author-published policy and use it directly
policy = AnyGuardrail.get_policy(GuardrailName.SHIELD_GEMMA, "dangerous_content")
guard = AnyGuardrail.create(GuardrailName.SHIELD_GEMMA, policy=policy)

See the Prompts & content guide. The catalogs are exported to schemas/guardrail_prompts.json and schemas/guardrail_content.json.

Full guides at docs link

Some of the models on HuggingFace require extra permissions to use. To do this, you'll need to create a HuggingFace profile and manually go through the permissions. Then, you'll need to download the HuggingFace Hub and login. One way to do this is:

pip install --upgrade huggingface_hub

hf auth login

More information can be found here: HuggingFace Hub

The guardrail space is ever growing. If there is a guardrail that you'd like us to support, please see our CONTRIBUTING.md for details.