Our AI is
bound to secrecy.

A professional bound by secrecy who copies a pleading, a file or a medical finding into a consumer service passes the content to a third party outside the client or treatment relationship. § 203 StGB makes exactly this disclosure a criminal offence, regardless of whether the provider later deletes or analyses the data. What matters is that the third party can gain knowledge of it at all.

A standard DPA covers only data protection; professional secrecy is a separate body of rules. Since the 2017 reform, § 203(3) StGB expressly allows professionals bound by secrecy to involve other persons in their work, provided these persons are bound to secrecy. Our DPA covers both: under its preamble we act as a person involved under § 203(3) StGB, and under section 6.5 of the DPA every person with access to protected information is bound to confidentiality in text form. The DPA applies from sign-up, so you don't need to sign anything separately. On request, you receive formal instruction on confidentiality under § 203 StGB and proof that our employees are bound to confidentiality.

No. If you delete names and leave out context, your answers are only half as good. With us, there is no need: confidentiality is governed by contract across the whole chain – with us, our employees, our data centres and the providers in the catalogue – and nothing is kept. If you still want to sanitise inputs before forwarding, switch on Noirdoc Shield as an option: our German infrastructure redacts details such as [PERSON] or [IBAN] statelessly and blocks secrets and prompt injections before the request reaches the model provider – configurable under Models → Shield as on demand, active by default or enforced.

Neither we nor the providers store your requests and answers or train on your data. If a provider deviates from this, we show it for each model, and you can block it per key. The log you keep as evidence contains timestamp, provider and model – only you know the content.

In three ways. First, confidentiality is agreed across the whole chain – with us, our employees, our data centres and the providers in the catalogue. If a provider lacks the agreement, we filter it out. Second, we keep nothing: no storage of your content, no training on your data. Third, inference runs at providers with data centres in Germany and the EU, and with Noirdoc you have a German contracting party.

Noirdoc Shield is an optional layer – under Models → Shield you set it to on demand, active by default or enforced, and you can override it per API key or request tag. When Shield is active, our German infrastructure inspects incoming text inputs before forwarding them to the model provider: personal and financial details are redacted with neutral placeholders such as [PERSON] or [IBAN], while secrets, custom blocked terms or prompt injections can be blocked outright. Inspection is stateless and one-way – we store no mapping state and do not re-insert original values into the answer.

No. Use for training is excluded in the contracts with the model providers, which we negotiated and pass on to you. We do not train any models on customer data ourselves – and nothing is stored after the answer anyway.

Open models such as Llama, Mistral, Qwen or DeepSeek at European providers, plus enterprise models through European hyperscaler regions. Which model suits you depends on the use case – long legal texts make different demands from short medical findings. Providers, models and their data residency are listed on the home page.

Usually not. The interface is OpenAI- and Anthropic-compatible: you swap the base_url and the API key, and everything else stays. Software that already talks to one of these APIs – from your firm's software to your own script – keeps working unchanged.

Through the policy under Models → Shield and your API keys. Across the organisation you set whether Noirdoc Shield runs on demand, active by default or enforced; for each key you define the permitted providers and models, the Shield mode and the spending limit. Every request is logged with timestamp, provider and model – that is your evidence for the professional chamber or the data protection officer. The log contains no content.

You pay for usage per model, billed by tokens, with no minimum spend and no setup project. The DPA including confidentiality under § 203 StGB costs nothing extra. Prices per model and provider are published on the home page.