Two calls. Your policies. In production.

Install the SDK (Python 3.9+), set your API key, and screen your first output. Keys are issued per environment at app.aetherlab.co.

pip install aetherlab
export AETHERLAB_API_KEY="your-api-key"

from aetherlab import AetherLabClient

client = AetherLabClient()  # reads AETHERLAB_API_KEY

# An AI response you want screened before it ships
ai_response = "Invest all your money in crypto! Guaranteed 10x returns!"

result = client.check_prompt(
    ai_response,
    blacklisted_keywords=["guaranteed returns", "financial advice"],
)

print(result.compliance_status)  # "Non-Compliant"
print(result.is_compliant)       # False
print(result.avg_threat_level)   # probability of policy violation
print(result.rationale)          # why it was flagged

At least one policy is required per check: configure standing policies in Policy Controls at app.aetherlab.co, or pass whitelisted_keywords / blacklisted_keywords per request.

import asyncio
from aetherlab import AsyncAetherLabClient

async def main():
    async with AsyncAetherLabClient() as client:
        result = await client.check_prompt(
            "Hello, how can I help you today?",
            blacklisted_keywords=["violence", "weapons"],
        )
        print(result.compliance_status)  # "Compliant"

asyncio.run(main())

All requests authenticate with an x-api-key header against https://api.aetherlab.co. The SDK reads AETHERLAB_API_KEY from the environment, or accepts api_key= explicitly.

• ·Never expose keys in client-side code; keep them in server environment variables.
• ·Use separate keys for development and production, and rotate them periodically.

Screens text against your guardrail policies and returns a compliance verdict with a threat level, confidence, and rationale. This endpoint backs the SDK's check_prompt().

* At least one policy is required: standing policies from Policy Controls, or a whitelist/blacklist sent with the request. Otherwise the API returns ERR_0202.

curl -X POST https://api.aetherlab.co/v1/guardrails/prompt \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "user_prompt": "Invest all your money in crypto! Guaranteed 10x returns!",
    "blacklisted_keyword": "guaranteed returns,financial advice"
  }'

The SDK method test_prompt() is a deprecated 0.3.x alias. Use check_prompt() in new integrations.

Screens an image against your guardrail policies. Accepts multipart form data with input_type set to file, url, or base64, plus the same policy and tuning fields as the prompt endpoint. Backs the SDK's check_media() and returns the same response shape.

result = client.check_media(
    "photo.png",
    input_type="file",
    blacklisted_keywords=["violence"],
)
print(result.compliance_status)

Both endpoints return a data object with the verdict. The SDK exposes it as ComplianceResult (compliance_status, is_compliant, avg_threat_level, confidence, rationale, full body as raw).

{
  "status": 200,
  "message": "Prompt Guard Response",
  "data": {
    "compliance_status": "Compliant",
    "avg_threat_level": 0.0,
    "confidence": 1.0,
    "rationale": "No violating content."
  }
}

{
  "status": 200,
  "message": "Prompt Guard Response",
  "data": {
    "compliance_status": "Non-Compliant",
    "avg_threat_level": 1.0,
    "confidence": 1.0,
    "rationale": "Guaranteed returns violate policy."
  }
}

Every check runs a multi-model compliance jury plus your policy engine: the content is evaluated by several models against your rules, and the verdict is adjudicated, scored, and explained. That is substantially more work per request than a single-classifier moderation ping, and the production medians reflect it: roughly 1.4s for text and 1.8s for images (median, across all reasoning modes).

• ·Budget ~2 seconds for a synchronous check in the request path; where your flow allows, run the check concurrently with other work.
• ·reasoning_mode controls the depth/latency trade-off per request. Talk to us about tuning it for latency-sensitive paths.
• ·The response includes the full verdict, threat level, and rationale in one round trip, so no follow-up call is needed to explain a block.

Standard HTTP status codes, plus an error_code field the SDK maps to a typed exception hierarchy (all inherit from AetherLabError).

Deploying at scale, on-prem, or with bespoke policies? Talk to us, or see how engagements work.