Three places it can run
Hosted by us
The application and the data sit on AWS in Frankfurt. Claude models run through Amazon Bedrock, in EU regions. There is nothing to install. Once you accept the offer, we set up your workspace with you.
In your own cloud account
The same application, installed in your AWS account, in an EU region you choose. Claude models run through Amazon Bedrock in that same account, so the bill for the models comes to you directly from AWS. We operate it with the access you grant, or your team takes it over.
On your own servers
Everything runs in your data centre, with an open-source model on your own hardware and no connection to the internet: the on-premise edition, version 2026.10.04. If you prefer Claude, the hybrid setup, available on request, keeps the documents with you and sends the model only the question and the relevant passages.
The three options compared
| Hosted by us | In your own cloud account | On your own servers | |
|---|---|---|---|
| Where documents and data are stored | AWS, Frankfurt, in our account | Your AWS account, in the EU | Your servers |
| Where the AI model runs | Claude, through Amazon Bedrock in the EU | Claude, through Amazon Bedrock in your account | An open-source model on your servers; on request, Claude through Amazon Bedrock in the EU (hybrid) |
| Internet access | Needed | Needed | Not needed with an open-source model |
| Who runs the servers and updates | We do | We do, or your team does | Your team, with our support |
| Getting started | An offer, then we set up your workspace | An offer, then the installation project | An offer, then the installation project |
Which model does the work
Claude, through Amazon Bedrock
The models we use today, chosen for their quality on Romanian documents and on long, multi-step work. They run only in the cloud and cannot be installed on your own servers. The closest to that is your own AWS account, in the EU. Bedrock does not use prompts or responses for training and does not pass them to Anthropic.
An open-source model, on your servers
Runs entirely on your hardware, with no internet connection. We choose the model with you, based on your tasks and your hardware, and test it on your documents before it goes into use.
Hybrid
On requestDocuments, search and the database stay on your servers. Only the question and the relevant passages go to Claude, through Amazon Bedrock in the EU, and nothing is stored there. You get Claude’s quality while the archive stays with you.
What stays the same wherever it runs
- The same application: library, conversations with sources, data tables, memory and automations.
- The same rule: nothing is sent, deleted or submitted without a person’s confirmation.
- Your data never trains a model.
- Roles for every user and a log of what the assistant did.
- A complete export of your data, on request.
What changes without internet
- The tenders radar, e-Factura and Google Drive need to reach SEAP, ANAF and Google. In a fully offline installation they stay off, or get access to those addresses only.
- The web assistant for your public website needs an address reachable from the internet.
- Scanned documents are read by a local text-recognition engine (Tesseract, Romanian and English).
- Updates arrive as new, tested versions of the application, installed by your team or by us with the access you grant.
For the IT team
- The application runs in containers: the web interface, the API, background processing, a task scheduler, a relational database, a search index, a message queue and file storage. The components are standard and need no separate licences.
- The on-premise edition runs on a single server. For larger volumes, we size the server with you (memory, cores, disk).
- Files live on the server’s disk.
- The local model runs on the server’s processor, behind an inference server with a standard API (llama.cpp); a GPU is optional and only makes it faster. We size the hardware with you, based on your document volume and the tasks you automate.
- Local accounts, with two-step sign-in (TOTP).
- Daily encrypted backups and the logs stay on your servers.
How we start
- 1
You tell us where the data has to stay and which rules apply: GDPR, internal policies, audits.
- 2
We send you a written offer: the option, the model, the scope, the schedule and the price.
- 3
Once you accept it, we test the chosen model on a few representative documents of yours.
- 4
We install, measure the quality on your documents and hand over, with documentation for your IT team.
Questions about private deployment
- Can Claude run on our own servers?
- No. Claude runs only in the cloud. The closest options, both on request, are your own AWS account in an EU region and the hybrid setup, where only the question and the relevant passages reach the model.
- Does it work completely offline?
- Yes, with an open-source model installed on your servers. Features that depend on outside services, such as the tenders radar, e-Factura or Google Drive, stay off or get access to those services only.
- Who installs the updates?
- We prepare and test them. Your team installs them, or we do, with the access you grant, on a schedule agreed together.
- How much does it cost?
- Each installation has its own written offer, based on the option and the volume, and you receive it before any work starts. In your own AWS account, you pay AWS directly for the models.
- Who is this for?
- Public institutions, companies with sensitive data such as health, legal or financial records, and anyone whose internal policy requires data to stay in their own infrastructure. For most companies, the version hosted by us in Frankfurt is simpler and cheaper.