Governance
Use GPT without sending customer data
Your support assistant runs on a US model, but names, emails and IBANs stay out. The answer comes back with the real values.
Comes back: The model's answer, with the real names back in place.
AI workflows for developers
Small models, chained.
One request.
Block a prompt injection, pull an order number from an email, strip the personal data, then hand the clean prompt to your model. One request, billed per character, EU-only.
Early access opens in batches. Every account starts with 1M characters free each month.
Example
- Customer messageinput
- Check for personal data and secretsdetect.pii.entitiesdetect.secrets.credentials
- Swap in placeholders, redact secretstransform.pii.tokenizetransform.secrets.redact
- Pass the message safely to big models like GPTrelay
# Needs jq. Set MESSAGE to the text to send.
jq -n --arg input "$MESSAGE" '{
"input": $input,
"detect": ["pii.entities", "secrets.credentials"],
"transform": [
{ "op": "pii.tokenize", "when": ["detect.pii.entities:flagged"] },
{ "op": "secrets.redact", "when": ["detect.secrets.credentials:flagged"] }
],
"relay": { "provider": "openai", "model": "gpt-4.1-mini" }
}' | \
curl -X POST https://api.chersus.com/v1/run \
-H "Authorization: Bearer $CHERSUS_KEY" \
-H "X-LLM-Key: $OPENAI_API_KEY" \
-H "Content-Type: application/json" \
-d @-Small AI jobs used to be hard.
| A big LLM | Your own small models | Chersus | |
|---|---|---|---|
| You pay | Tokens, for every job | GPUs, busy or idle | Characters in and out |
| Speed | A general model thinks it over | Fast, once warm | Super fast, always warm |
| Who sees your text | The model provider | You | Chersus, in the EU, keeping none of it |
| Who operates it | The provider's staff | Your people, on call | We do |
Now they fit in one request.
All twelve examples →Security
Check content before your agent reads it
Web pages, emails and documents can hide instructions for your agent. They are blocked before the text reaches the prompt.
Comes back: A block, before the page reaches your agent's prompt.
Data
Keep personal data out of your logs
Log lines and tickets reach your warehouse or monitoring tool without names, emails or keys.
Comes back: The same log line, minus names, emails and keys.
Content
Sort a multilingual support inbox
Tickets reach the right team by language and topic, without keyword rules, and without personal data in storage.
Comes back: The language and the topics, ready to send to the right team.
Your model can be part of it.
input
detect
extract
transform
relayrehydrate
output
- 01Send your text to
/v1/run, with the steps you want. - 02Detects and extracts read it at the same time. Transforms rewrite it, in the order you list them.
- 03Relay sends the clean prompt to your model and puts the real values back. Or take the clean text and call your model yourself.
Nothing leaves your browser.
Sixteen operations to combine.
All operations →detect7
A verdict and a score. Personal data, credentials, prompt injection, jailbreaks, hate, abuse and threats.
- pii.entities
- toxicity.hate
- toxicity.abuse
- toxicity.threat
- prompt.injection
- prompt.jailbreak
- secrets.credentials
extract6
A value. Language, fields, title, description, topics and emoji.
- fields.schema
- language.code
- content.title
- content.description
- content.topics
- content.emoji
transform3
Rewritten text. Redact or tokenize personal data, redact credentials.
- pii.redact
- pii.tokenize
- secrets.redact
relay
The model call. Your model gets the clean prompt; your user gets the real names back.
Read in Europe, then gone.
European models, running on EU infrastructure in 26 cities. Chersus processes your text there and keeps none of it. With your own key, the provider you relay to is your processor, not ours.
- European models, processed in the EU
- A European stack, no US hyperscalers
- No content kept
- A hash-only audit trace on every response
Pay per character, not per token.
Calculate your bill →- Free tier
- €01M characters free, every month
- Starting rate
- €0.06per 1K characters, to start
- Volume
- −20%with every doubling
Combine any operations. You pay for the text you send and the text you get back.
What people ask first.
How fast is it?
Every operation runs on a small model built for one job, not a general model that thinks it over. Try it on your own traffic in early access.
Does the model ever see personal data?
Only what you let through. A tokenize step swaps personal data for placeholders before the prompt leaves, and relay puts the real values back in the answer. Detection is statistical, so treat it as a strong filter, not a guarantee.
Which languages?
It depends on the operation: personal data in 27, language detection in 84, topics in 101. Each operation lists its own.
Do I have to use relay?
No. Take the clean text and the token map, call your model yourself, and put the values back. Relay does all of that in one request.
Where does my data go, and for how long?
Chersus processes it in the EU and keeps no content after the request. The audit trace holds hashes only. With relay, the prompt also goes to the model you named.
Your first request is one curl away.
Access opens in batches. Request yours, get your key, and start with 1M characters free every month.