How long does the first run take?

Minutes. Start the runtime, post a prompt-injection string at the evaluate endpoint and watch it get blocked, post a string with an email address in it and watch it come back redacted, then wrap one real model call with the SDK. That sequence is the whole quickstart.

Do I need the hosted console to try it?

No. The open-source runtime reads a local YAML policy file and enforces without a control plane. Add the console later when you want shared policy, version history, traces, and audit evidence across a team.

How do I turn on ML detection?

Point the runtime at an in-boundary inference sidecar and pin the model weights you want it to serve. Until you do, the engine runs its deterministic layer only - which is a real security control, not a demo mode.

Getting started

Enforcing policy in two steps, without an account.

The runtime is open source and runs standalone, no signup, no telemetry, no limits. Add the hosted control plane later, when you want centralised policies and team workflows.

Before you start

Docker, or a compatible container runtime, on the host that will run it. Install Docker →
Pull the images. Optional - each docker run pulls what it needs if it is not already on the host. The inference image is only for the ML tier.

docker pull znyx/runtime:latest

# only for the ML detectors - see the optional section in step 1
docker pull znyx/inference:latest

Start the runtime

Run this beside your app. Local mode is the default, so there is no account, no key, and no outbound calls - the runtime enforces its deterministic rules on its own.

docker run -p 8080:8080 znyx/runtime:latest

Mount your own policy with -v $(pwd)/policies.yaml:/app/config/policies.yaml

Install the SDK

Add the SDK to your project. Six first-party SDKs wrap the same evaluation contract, or call the HTTP API directly.

pip install znyx-sdk

TypeScript: npm install @znyx/sdk

Integrate

Six SDKs, one contract

Python · POST /v1/evaluate/input

fail-closed

from znyx_sdk import GuardrailsSyncClient

guardrails = GuardrailsSyncClient("http://localhost:8080")

input_result = guardrails.evaluate_input(
    user_message,
    tenant_id="default",
    app_id="my-app",
)
if input_result.is_blocked:
    return {"error": input_result.user_message}

llm_response = call_llm(input_result.sanitized_text or user_message)

output_result = guardrails.evaluate_output(
    llm_response,
    tenant_id="default",
    app_id="my-app",
)

return output_result.sanitized_text or llm_response

Before production

Production readiness checklist

Six things worth confirming before the firewall carries real traffic. Each maps to a check in the CLI.

Full deployment guide →

Runtime responds 200 on /healthz inside your VPC
Inference sidecar reachable for ML and judge tiers
Policy bundle pinned to a version, not to latest
Fail-closed behaviour confirmed under sidecar outage
Output / DLP gate enabled on every user-facing path
Traces exporting to your OpenTelemetry collector

FAQ

Getting started questions

What you need, how long it takes, and what to try first.

Docker, or Python if you would rather pip install the runtime. No account, no API key, and no GPU. The runtime is rules-only out of the box, which is what makes the first run a single command.

Enforcing policy in two steps, without an account.

Before you start

Start the runtime

Optional: add ML detection with the inference sidecar+

Install the SDK

Six SDKs, one contract

Production readiness checklist

Getting started questions

What do I need before I start?−

How long does the first run take?+

Do I need the hosted console to try it?+

How do I turn on ML detection?+

Secure every prompt, agent, and tool call, in your boundary.