AI Gateway
The Cervus Gateway sits between your applications and the language models you use: it controls access, masks personal data, enforces budgets and records what was used. Operated in a data centre in Germany, with an OpenAI-compatible interface. Included in our managed local AI packages (page in German).
Without a central layer, every application holds its own provider keys and its own rules. Nobody can say who used which model for what, personal data travels in prompts unchecked, and a faulty automation is only noticed on the invoice. A gateway is the one place where those questions are answered.
Access
API keys are tied to versioned use cases, and each key is cleared for specific models. On request, stating a use case can be made mandatory. Organisations can be split into departments, with separate roles for administration and use.
Personal data
An optional module detects personal data in prompts and replaces it with placeholders before the request reaches an external model. Detection is tuned for German formats. It reduces risk – it does not remove the need for rules and training.
Cost
Per key: requests per minute, tokens per request, parallel requests. Per organisation: token or euro budgets with a hard ceiling. A runaway automation stops at the limit instead of running up a bill.
Evidence
Every request is logged with time, key, model, route, token usage, cost, latency, status, use case and trace ID. Prompts and responses are not stored by default. Settings changes land in an audit log; exports are available as CSV or JSON.
Where it runs
We run the gateway in a data centre in Germany. Behind it you can use cloud models from providers such as OpenAI, Anthropic, Google or Mistral, managed open-weight models from European data centres, or your own on-premise models – or any combination of these.
Which providers and locations are involved is documented in the project and covered by a data processing agreement. Where a request is processed stays your decision, per key and per use case.
Integration
The interface is OpenAI-compatible: chat completions including streaming, the responses API, embeddings and the model list. Tools that let you set a base URL – n8n, your own software – need a base URL, an API key and a model name.
Which team may use which model, for which documented purpose, with which rate limits and budgets. Masking and prompt-injection checks can be switched on per organisation or key.
Metadata logs show which model is actually used, what it costs and where requests fail. Exports give you the evidence for internal reviews.
Not supported at the moment: image and audio endpoints and the older completions interface. Prompts should be tested against the model you choose.
Questions we get
Not by default. The gateway records metadata about each request. Storing content is optional per organisation or key – masked or complete – with a retention period you set. Metadata logs are kept for 365 days by default.
Yes. The gateway works on its own in front of cloud models. Local models can be added later without changing your applications.
No product does that on its own. The gateway provides technical controls and evidence that support your governance – the assessment of your specific use case stays with you and your legal advisors.
German and English. Our contact form is in German; you are equally welcome to write to info@cervus-it.de or call +49 991 40545335.
We look at your use cases and set up the gateway accordingly – on its own or together with local models.
Get in touchDetailed German pages: Cervus Gateway · What is an AI gateway?
