A pretrained toxic language classifier that sorts text into one of 2 categories. Use the toxic language API immediately, no training required, then adapt it to your own data when you need more.
Drop in some text and get the prediction back. No signup, no setup.
A sample of the 2 labels this pretrained classifier chooses between.
Need a label that isn't here? Clone the classifier into your Nyckel console and edit the label set to fit your data.
Get your own copy of this classifier behind your own endpoint — callable from any HTTP client:
curl -X POST "https://www.nyckel.com/v1/functions/YOUR_FUNCTION_ID/invoke" \
-H "Authorization: Bearer YOUR_ACCESS_TOKEN" \
-H "Content-Type: application/json" \
-d '{"data": "The text you want to classify"}'import requests
response = requests.post(
"https://www.nyckel.com/v1/functions/YOUR_FUNCTION_ID/invoke",
headers={"Authorization": "Bearer YOUR_ACCESS_TOKEN"},
json={"data": "The text you want to classify"},
)
print(response.json())const response = await fetch("https://www.nyckel.com/v1/functions/YOUR_FUNCTION_ID/invoke", {
method: "POST",
headers: {
"Authorization": "Bearer YOUR_ACCESS_TOKEN",
"Content-Type": "application/json",
},
body: JSON.stringify({ data: "The text you want to classify" }),
});
console.log(await response.json());$ch = curl_init();
curl_setopt($ch, CURLOPT_URL, 'https://www.nyckel.com/v1/functions/YOUR_FUNCTION_ID/invoke');
curl_setopt($ch, CURLOPT_RETURNTRANSFER, 1);
curl_setopt($ch, CURLOPT_POST, 1);
curl_setopt($ch, CURLOPT_POSTFIELDS, '{"data": "The text you want to classify"}');
$headers = array();
$headers[] = 'Authorization: Bearer YOUR_ACCESS_TOKEN';
$headers[] = 'Content-Type: application/json';
curl_setopt($ch, CURLOPT_HTTPHEADER, $headers);
$result = curl_exec($ch);
curl_close($ch);
echo $result;Example response
{
"labelName": "Toxic",
"labelId": "label_...",
"confidence": 0.92
}Trained on a Nyckel-curated dataset covering 2 toxic language categories, served on Nyckel's own infrastructure — your text snippet stays on Nyckel.
Send raw text to the invoke endpoint; the response is a label with a confidence score.
Clone it, then correct predictions and add your own samples in the console — Nyckel retrains automatically, turning this into a custom model tuned to your data.
Identify and remove offensive content to maintain a safe online environment.
Filter out toxic comments in real-time to improve the support environment for agents and customers.
Enhance player experience by detecting and acting on toxic behavior among participants.
Protect students from harmful content in forums and discussion boards.
Provide moderation solutions to forums and websites to maintain community standards.
Screen for inappropriate language in job applications or internal communications.
Automate the detection of offensive content to ensure constructive community discussions.
A zero-shot classifier uses a large foundation model's general knowledge to pick between your labels — no task-specific training, so new or edited labels work immediately. A Nyckel-trained classifier has been trained on labeled examples and runs on Nyckel's own infrastructure, which typically makes it faster, cheaper per call, and more accurate on data that resembles its training set. The "Under the hood" section on this page shows which kind this classifier is, and any classifier can be adapted into a trained one by adding your own examples.
Honestly: we can't know in advance — it depends on your data stream and how closely it resembles what this classifier has seen. The reliable way to find out is to measure it on your own data: start invoking the classifier with real traffic, or upload and annotate a set of text samples in the console — make sure they look like your production data, not idealized examples. Nyckel's evaluation metrics then show you exactly how it performs on that data before you rely on it.
No classifier is perfect, so Nyckel is built around the correction loop: invokes can be captured for review, you confirm or correct predictions in the console, and corrections become training data. Over time the model adapts to your data distribution — accuracy on your traffic improves with use rather than staying fixed.
No. This toxic language classifier works out of the box — clone it into your console and you'll have your own API endpoint in under a minute. Training data only enters the picture when you want to adapt it: your corrected predictions and uploaded samples improve the model, and you can also edit the label set to match your needs.
Trying the classifier on this page is free with no signup. Cloning it requires a free account, and the free tier covers your first API calls each month — see nyckel.com/pricing for current limits and paid tiers.
Add this pretrained classifier to your Nyckel console — you'll get a live API endpoint in under a minute, and a path to a custom model when you need one.