Smaller models than the US labs
We run strong open-weight models on European hardware. They're not the biggest proprietary models — that's the price of keeping the whole chain in the EU.
The long version: every part of the workspace, how it compares to US assistants and other EU platforms, and exactly which infrastructure runs it.
Mitarbot fits teams with real documents, real recurring work and a reason to keep data in Europe.
Pick a provider and see the comparison. Prices and product facts checked against public information in October 2026; competitors change fast, so verify before you buy.
Choose a provider above to see how Mitarbot compares.
Sovereignty has a price. We'd rather say it up front.
We run strong open-weight models on European hardware. They're not the biggest proprietary models — that's the price of keeping the whole chain in the EU.
No mobile apps, no SSO/SAML yet, fewer connectors than the big platforms. The core — files, chat, mini-apps, schedules, FTP, Drive, email — works today.
Web search sends only your query to search providers, image generation only the prompt outside the EU. Your files, chats and customer data never do.
No SSO/SAML, SCIM, org-wide admin or certifications yet. Everything lives per workspace — members, files, connectors and budget — with no company-wide layer above it.
You pay for the tokens the agent uses and for the time the workspace is running. Nothing else — and nothing while it sleeps.
There is no subscription, no seat price and no minimum: top up a balance, use the workspace, and a month of not using it costs nothing. Exact rates are shown in the app before you top up.
Billing in the app is an estimate for budgeting; the balance is the hard cap. When it reaches zero, the run stops.
Every component, by name, with its jurisdiction.
A DPA and the current sub-processor list are available on request. Data is deleted with the workspace; snapshots are removed with it.
Fine print: optional web search sends only your search query to third-party search APIs, and image generation currently sends only the image prompt to an international endpoint. Everything stored stays in the EU either way.
On a Hetzner server (fsn1) in Falkenstein, Germany, in a volume dedicated to your workspace. Snapshots go to Hetzner Object Storage in the same location. There is no US replication and no US cloud involved.
The European gateway Cortecs routes to TensorX, an Irish provider running inference on EU-sovereign GPUs in Dublin and Helsinki. The chat model is GLM 5.3 Flash; images are described by MiniMax M3. TensorX keeps zero data and does not train on prompts.
No. Your files, chats and mini-apps are never used for training, and inference providers keep no retention of your prompts.
Sandboxes suspend after 15 idle minutes and cost nothing while asleep. A month without activity costs €0. After a year of inactivity we archive the workspace and offer you a download.
Yes. A workspace exports as a ZIP with chats, files, jobs and connectors, and deleting a workspace removes its database rows, container, volume and snapshot.
Because the US labs' EU processing is limited to certain enterprise tiers and usually covers stored data rather than inference. If the requirement is 'no US company processes our work', the only reliable way is to avoid those models entirely — which is what Mitarbot does.
The live demo is the real product with prepared data — read-only, no account, both languages.