Sensitive data stays in Europe
Use customer, client and employee data in AI tools without handing processing over to providers subject to the US CLOUD Act. The core and inference run with European providers without US parent companies.
Souveräne Systeme
AI services on European infrastructure
Souveräne Systeme is a platform where your business rents individual AI services — chat, research, documents, conversation translation, email triage and more. The processing services run with European providers without US parent companies. Content is available in plain text during processing. Saved results use zero-knowledge encryption by default; plain-text storage requires explicit opt-in and is unavailable for § 203 accounts.
Benefits for your business
AI tools are now part of everyday work — but customer, client and employee data do not belong in US cloud services. The Souveräne Systeme platform addresses this gap.
One account, several tools
Your team uses chat, research, document preparation, conversation translation, email triage and the other services through one account. Employees receive their own accounts with their own keys and share the owner's balance. Each service is billed by usage — instead of several separate subscriptions with unclear costs.
A verifiable legal basis
Data processing agreements under Art. 28 DSGVO with every subprocessor; we document the chain openly. A separate processing route for professionals bound by secrecy under § 203 StGB has been designed but is not yet operating.
The services
Each service is a separate module with its own data policy. You book only what you need. All services listed here are operating — work in progress is not listed on this page.
Chat
AI chat using current open-weight models, which you can switch between during a conversation. With projects, a prompt library, file extraction and access to your own knowledge base — for lookups and, after confirmation, saving entries too. For calculations and spreadsheet analysis, the assistant writes Python and runs it in an isolated environment in your browser instead of generating the number itself: the spreadsheet stays on your device for this, and the calculation steps appear below the answer.
Workflows
A fixed sequence of one to five steps from a closed tool catalog, which you create once and then start in chat. The twelfth run follows the same steps as the first; if the result is not useful, change the step instead of your wording.
Research
A research assignment, search queries shown to you verbatim for approval before they are sent, and a report at the end with numbered sources and an as-of date. Optional web tools in Chat, Workflows and Site Watch can also send queries or fetch pages beyond our infrastructure, subject to their consent and safety checks; web access is blocked for § 203 accounts. The transparency report describes these routes.
Documents
Read PDFs, Office files and scans (with text recognition), detect personal details and redact them before sharing. Preparation before a document goes into an AI tool. The file can also come from your own Nextcloud instead of being moved to us.
Klangskript
Transcribe voice recordings (file upload) and show spoken words live as translated captions — one person speaks, listeners read in their own language, even without an account in briefing mode. Processing runs exclusively in the memory of European GPU instances.
Own instance
A GPU instance exclusively for your account for the duration of the session, with an open model of your choice instead of the shared inference API used by chat. Billing is based on runtime. The instance is not approved for accounts under § 203 StGB: it lacks the written contractual clause with the data center, and exclusive use does not replace that clause.
Terminal
Your own computer's command line in a panel along the bottom, intended for working with a command-line tool connected to our API. The browser connects to 127.0.0.1, meaning your computer; keystrokes and output do not pass through our servers.
Open Notebook
Research notebook: collect sources, ask questions about them and, if you wish, summarize them as an audio podcast. Podcast voices are generated on self-hosted systems within the European session perimeter; content never leaves it.
Postvogt
Email triage: sorts incoming messages into configured folders and creates draft replies where appropriate. Sending is always done by a person, without exception.
Matrix bot
The assistant in your Matrix rooms. We do not operate a homeserver — you create a bot user and can revoke its access yourself at any time. Only users you have approved are allowed. Technically, it works on any homeserver; we strongly recommend your own, because encryption protects message text but leaves membership, timestamps and room names visible to the server operator.
Calendar
Your existing calendar over CalDAV, read at the moment you request it. We do not operate a calendar of our own, and we do not store any appointments. Anything the assistant would create, change or delete is presented to you for confirmation first.
Deadline Watch
Record contracts — by upload with suggested key details and a quote from the clause, or manually — and receive an email reminder before the cancellation deadline. The reminder contains the title and date, never the contract contents.
The data boundary
Our architecture separates the account and billing core from content processing. Saved results use zero-knowledge encryption by default; content is available in plain text during processing.
Live content bypasses the core
Our core system (Hetzner, Germany) processes accounts, billing and encrypted data blocks. Live and raw data never reach it: your browser sends them directly to the processing layer, which holds them only in memory and does not store them. Saved results are held in the core — encrypted by default, or as plain text only if you explicitly choose it for that category. This opt-in is unavailable for § 203 accounts.
Histories encrypted on your device
By default, saved chat histories are encrypted in your browser with a random account key that never leaves your device. A password-derived key only wraps that account key; it does not encrypt each history directly. The core cannot read these encrypted histories. Plain-text storage requires explicit opt-in and is unavailable for § 203 accounts. A password reset cannot restore zero-knowledge histories — you receive recovery codes when you register.
European providers, openly documented
The core system runs at Hetzner (Germany), which also stores the encrypted files in your knowledge base. Since 30 August 2026, the always-on services — Chat, Documents, Matrix bot, Site Watch — have run on a dedicated content-processing server at netcup (Nuremberg). A separate netcup server in Germany forwards optional web searches and fetches pages. GPU sessions run at Scaleway (France/Poland); since 21 September 2026 the chat model API has run at AKI.IO (Berlin, data centres exclusively in Germany) — each covered by a data processing agreement under Art. 28 DSGVO. A separate route for professionals bound by secrecy under § 203 StGB has been designed but is not yet operating. None of the providers we commission is subject to the US CLOUD Act as a US company. For web search, you may bring your own provider — including an American one; that is your choice and is off by default. We document our subprocessor chain openly.
Storage modes
For each content category — chat histories, documents, conversation results, research reports — you decide how it is retained. Zero-knowledge encryption is the default.
Ephemeral
Nothing is saved beyond the session. Content exists only in memory for the duration of processing and is overwritten afterward.
Zero-knowledge (default)
Saved results are encrypted in your browser with a random account key that never leaves your device. A password-derived key only wraps that account key; it does not encrypt each result directly. In this storage mode, the core stores only ciphertext; processing services still handle content in plain text.
Plain text (opt-in)
An explicit choice for categories where you do not want to accept the trade-off of encryption: without your password and recovery code, zero-knowledge content is permanently lost, even to us. The core does not analyze or search saved plain-text content. This mode is blocked on the server for accounts under § 203 StGB.
At launch
The platform is expected to launch in mid-November 2026. Before that, we have our contracts and data protection reviewed by a lawyer and grant no new access. Email us and we will get back to you once, when we launch.