Integration guide

The integration model is simple: evaluate before and after the model call.

The API surface does not change when you move from Starter to Growth. The runtime keeps the same endpoints; Growth and Enterprise change how policies are operated, not how your app calls the runtime.

Request flow

User -> Your App -> /v1/evaluate/input -> Model -> /v1/evaluate/output -> User

For tool calls, insert /v1/evaluate/tool before execution. For streamed output, use /v1/evaluate/stream while tokens are still arriving.

Runtime endpoints and where each one fits in the application flow.
Endpoint	When to call it	Why it exists
POST /v1/evaluate/input	Before the model call	Check prompts or tool inputs before they reach the model.
POST /v1/evaluate/output	After the model call	Filter, block, or transform model output before returning it.
POST /v1/evaluate/tool	Before tool execution	Evaluate tool requests and arguments before execution.
POST /v1/evaluate/stream	During streamed output	Apply runtime checks while a response is still being generated.

Core request fields for scoped policy evaluation.
Field	Required	Purpose
text	Yes	The prompt, output, or tool payload to evaluate.
tenant_id	Yes	Tenant or organization identifier for scoped policy resolution.
app_id	Yes	Application identifier so multiple apps can share the same runtime safely.
agent_id	No	Optional agent or workflow scope.
env	No	Optional environment such as dev, staging, or prod.
metadata	No	Optional structured context for policy and observability use cases.

Python example

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

TypeScript example

import { GuardrailsClient } from '@znyx/sdk';

const guardrails = new GuardrailsClient('http://localhost:8080');

const inputResult = await guardrails.evaluateInput({
  text: userMessage,
  tenantId: 'default',
  appId: 'my-app',
});

if (inputResult.isBlocked) {
  return { error: inputResult.userMessage };
}

const outputResult = await guardrails.evaluateOutput({
  text: llmResponse,
  tenantId: 'default',
  appId: 'my-app',
});

return outputResult.sanitizedText ?? llmResponse;

Operating note

Starter works with local policy files. Growth connects the same runtime to a hosted console. Enterprise can self-host the console as well. Your integration stays stable across all three.

Architecture Compare Plans