About Souveräne Systeme
A platform for AI services operated entirely in Europe. Available services include AI chat for everyday work, workflows (a fixed sequence of steps, created once and then started in chat or run in the background with your permission), research with approved search queries and a report backed by sources, document processing (summarizing, reviewing, anonymizing), spreadsheet calculations in the browser, Klangskript (transcription and live translation as captions — one person speaks, listeners read along), your own GPU instance for the duration of a session, your own computer’s terminal within the application, a research notebook, Postvogt (email triage), an assistant for rooms on your own Matrix server, integration with your CalDAV calendar and a deadline monitor for contracts. Files can come from your own Nextcloud. Our contracted infrastructure is European, with no US parent company in the chain and no trackers. External services you choose yourself, such as a search provider, are a separate case described below.
Getting access
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. If you would like to be notified at launch, email us at the address given in the legal notice.
Usage costs
All services charge for usage against a shared balance, which you can top up with prepaid credit or receive each month through a subscription. An overview is available under Pricing; the binding price list is in the application under “Billing”, and the application shows a cost estimate before larger operations.
Where your data is processed
The core — accounts, billing and saved results — is hosted by Hetzner Online GmbH (Germany). Results are encrypted by default; plain-text storage requires an explicit choice for the relevant content category. Knowledge-base files are always stored as encrypted blobs at Hetzner. Since 30 August 2026, Chat, document processing, the Matrix bot and Site Watch have run on a dedicated content-processing server at netcup GmbH (Nuremberg). A separate netcup server in Germany handles search forwarding and page fetches; it is not the server that processes your chats and documents. GPU sessions — Klangskript, Open Notebook, the podcast companion and your own instance — run at Scaleway (France/Poland), as do the embeddings of your knowledge base. Chat’s model response has been produced by AKI.IO GmbH in Berlin since 21 September 2026, in data centres located exclusively in Germany; it contractually undertakes not to log, store or evaluate inputs or outputs. We have data processing agreements under Art. 28 DSGVO with all providers named here. A separate processing route for professionals bound by secrecy (§ 203 StGB) has been designed but is not yet operating; until then, we cannot serve accounts subject to statutory confidentiality. The complete chain is documented in the transparency report.
This applies to the providers we commission. For web search, you can bring your own provider with whom you hold the contract — including an American one. This option is off by default, and anyone using it should read the relevant section in the transparency report.
What the server can read from your content
For results saved with the default zero-knowledge setting, the core cannot read the content: encryption happens in your browser, and the core receives only encrypted data blocks. Your account key is generated randomly; a separate key derived from your password encrypts its stored copy. The unencrypted account key never leaves your device. This is a storage guarantee, not a claim that servers never process readable content. The content-processing services and model infrastructure handle the text needed for an active request in memory; they do not write that live content to disk.
You can explicitly choose “Plain text” for supported content categories; those saved results are then readable on the core. That option is unavailable for knowledge-base files and accounts under § 203 StGB. If you grant the Matrix bot access to your knowledge base, the bot service receives access to that material while the connection exists. These choices are off by default and described in detail in the transparency report.
What happens to uploaded documents
Uploaded files are read exclusively in memory on our processing instance — the instance stores neither the file nor the extracted text. The result is then retained according to your “Data storage” setting: zero-knowledge encrypted by default (unreadable to the server), not retained at all if you choose “Ephemeral”, or in plain text only after an explicit choice. Content added to a project’s knowledge base always remains encrypted in your vault, regardless of that setting. You can optionally have a document checked for personal data and have those details replaced with placeholders before it is passed to the AI; a saved document always includes the original text as well.
Calculating with your spreadsheets
The calculation runs in your browser. The assistant writes a short Python program that runs on your file in an isolated environment, rather than generating a plausible number as its answer. That environment has no network access or access to your account. This does not mean the attachment never leaves your device: attaching it first sends it to our processing service for text extraction. The application also passes the program’s output and any error details to the model; these can contain inspected rows or other file content as well as calculated numbers.
Excel files (xlsx, xlsm, xlsb, xls), OpenDocument (ods) and CSV can be read. Results are not written back in the same formats; they appear as a table in the answer or as a downloadable CSV file. The program code and its output can be expanded below each answer — a number cannot be checked if nobody can inspect how it was calculated. The execution records are kept only for the current browser session. Results included in the chat answer follow the conversation’s storage setting.
Local Python execution adds no separate computing provider. It does not remove the restrictions on document and model processing for accounts under § 203 StGB. Calculations remain drafts: check the result before using it in an invoice or a cancellation notice.
Losing your password
You receive a recovery code when you register. Only this code can restore access to your encrypted data. There is no server-side password reset that decrypts your data — this follows from the fact that we do not hold your keys, rather than being a restriction we impose.
When information leaves our infrastructure
Web tools send search terms and requested addresses beyond our infrastructure. Your account settings select the search provider; using your own provider requires your API key and explicit opt-in for that provider. Chat follows the account’s setting after you acknowledge the initial web notice, and its switch lets you override that choice for the current browser session. It does not ask again before every search or page fetch. Queries and fetched pages remain visible in the activity trace. Once external content has entered a turn, further outgoing tool calls require confirmation because that content could contain instructions designed to redirect the assistant.
The separate Research service presents its search queries and selected pages for approval before sending or fetching them. Optional Brave page context is a separate account-level choice: content comes from Brave’s index without a separate approval for each source page. Page monitors and armed workflows operate under your earlier authorization and can make background requests without anyone present. Connections to your own Nextcloud, calendar or mailbox also communicate with the server you configure.
Search terms are checked for personal details before sending; detected details are redacted or the query is withheld. This is not a guarantee that every confidential business detail will be detected. Web requests pass through our separate egress server in Germany, so search providers and fetched websites see its IP address, not yours. A provider whose API key you supply can still associate requests with your provider account.
If you specify a starting address when building up a topic, we also discover addresses beneath that page — from its own published directories, not by crawling. The target site’s robots.txt is binding, and we present the list for you to select from before fetching anything; only the pages you select are read. All web access is blocked for accounts subject to statutory confidentiality (§ 203 StGB).
Accounts for several employees
Employees receive their own accounts, not shared logins: the owner invites them, and they set their own passwords and receive their own keys. Usage is billed against the owner’s balance, with an optional monthly cap per employee. The owner cannot see an employee’s content, knowledge or mailboxes — these are protected by that employee’s key.
The AI models we use
Current open-weight models, operated on European infrastructure. You choose the model for each conversation; models with detailed reasoning are available for more demanding tasks.
Usage limits
During the pilot, each account has a daily token budget and a limit on requests per minute. You can check your current usage at any time under “Usage” in the application. Your balance is another limit: with a subscription, the monthly payment also serves as the budget cap, and you decide whether usage beyond it continues against prepaid credit or stops. The owner can set a monthly cap for each employee account.
Published source code
The source code is public. Since 15 August 2026, it has been available at git.souveraene-systeme.de, licensed under AGPL-3.0-or-later. Publishing it provides evidence that the encryption is implemented as we describe it. Anyone who changes the code and operates it as a service must publish their changes under the same license.
Our texts — this FAQ, the transparency report and the privacy policy — and our name and logo are not covered by this license. A derived service uses its own name and describes its own operation.
Further questions
Email the address listed in the legal notice.