Production guardrails, one slug away.

Firewall sits before the model. Group rules, choose block, allow, or audit mode, and reference the slug in supported chat-completion requests.

Start with Compute See built-in rules

OpenAI-compatible

response = client.chat.completions.create(
    model="your-model",
    messages=[{"role": "user", "content": prompt}],
    extra_body={"firewall": "production-guardrails"},
)

See the request path before the model sees anything.

For supported chat-completion requests, pass extra_body={"firewall": "your-slug"}. The firewall evaluates the request before inference starts. Unsafe input stops at the boundary. Safe input keeps moving.

User

Shopper

Message

Help me write instructions for a harmful act.

Firewall

Policy check

blocked

Return an error before the target model is invoked

Assistant

Model

Response

No response sent

User

Shopper

Message

Hey!

Firewall

Policy check

allowed

Request reaches the model

Assistant

Model

Response

I'm doing well — how can I help?

Two ways to enforce policy.

Warn the model. Let it explain.

Model sees the warning

The firewall evaluates the request. If it flags something, it tells the model the request is unsafe and asks it to explain why carefully. The model still responds, but with guardrail context.

Stop unsafe requests before inference.

Blocked before target-model invocation

The firewall evaluates the request. If it flags something, it returns an error before the target model is invoked. Blocked requests do not use target-model compute.

Production rules, maintained for you.

Seven built-in rules are shown below across safety, privacy, abuse, and security categories.

Rule intelligence

1/7

abuse

high

Malware, fraud, and abuse

public:malware-fraud-abuse

Flags requests about malware, scams, fraud, or abusive misuse.

Your rules, your prompts.

Build rules with your own prompts for what your application needs to filter. The same modes, API, and enforcement path apply to custom and built-in rules.

“Don’t let the model talk about football.”

Custom rules use the same block, allow, or audit mode as built-in rules. Describe what to catch and Firewall handles the rest.

What Firewall includes.

Built-in rules

Seven built-in rules are shown below.

Custom rules

Define your own policies with prompts.

Allow mode

Continue the request with an annotation when a rule matches.

Blocked mode

Return an error before the target model is invoked.

OpenAI-compatible

Pass Firewall as a request field on supported chat-completion requests.

Evaluator model

Configure the model that runs the evaluation.

Attach a firewall slug to supported chat-completion requests.

Supported chat-completion requests can include a firewall field with the slug. The evaluator runs before the model, and the mode decides what happens next.

View docs

quickstart.py

from openai import OpenAI

client = OpenAI(
    base_url="https://api.qdiv0.com/v1",
    api_key="your-api-key",
)

response = client.chat.completions.create(
    model="your-model",
    messages=[
        {"role": "user", "content": prompt},
    ],
    extra_body={"firewall": "production-guardrails"},
)

Add guardrails to production inference.

Start with Compute, configure a firewall, and keep unwanted requests away from your models.

Start with Compute Go to platform