Changelog

This release makes previewing generated files more reliable and removes a manual step when you create an agent.

Improvements​

• Create Agent: When you attach knowledge to an agent, Zentre now selects a model and mode that supports knowledge automatically.

Fixes​

• PDF preview: Opening a PDF after previewing a Markdown file no longer shows an "Invalid PDF structure" error.
• Chat preview: You can open another generated file while a CSV preview is open.

New workspaces now start with a redesigned onboarding journey. The new screens guide you through the first steps in Zentre and follow our updated design system.

Improvements​

• PII masking: Masking of personal data in chat has been validated for English and German. When PII masking is switched on and the masking service cannot process a message, Zentre does not send the message to the model instead of sending it unmasked.

You no longer have to choose a model for every question. In auto mode, Zentre picks the model, the thinking budget, and whether web search is available for each message.

Auto mode only uses models your team has access to. Manual model selection works as before.

Fixes​

• Scanned PDFs: Content from scanned PDFs is now extracted and can be found through knowledge questions.

Start from a template instead of a blank page. The new template catalogue lets you browse ready-made agents and create your own agent from one in a few clicks.

Templates and categories​

Templates are grouped into categories, so it is easier to find an agent for a specific task. Templates built for §203 Professional Secrecy Mode are only visible to organisations entitled to use that mode.

If a template uses connectors, Zentre matches them with the connectors your organisation has already set up.

Conversation starters​

Agents can now offer up to five conversation starters. They appear in chat and help users understand what an agent is good at.

Improvements​

• Prompt enhancement: "Enhance" now produces structured agent instructions, and longer instructions are no longer cut off.
• Chat: Template tabs have been consolidated, connector names are shown, and the knowledge tabs have clearer labels.

You can now create images directly in a conversation. Ask for an image in chat and Zentre generates it for you. Generated images stay in the conversation when you reload the page.

Each image is charged to your organisation's balance. Image generation is not offered inside §203 conversations.

Zentre now offers a dedicated mode for professions bound by professional secrecy under Section 203 of the German Criminal Code (StGB).

How it works​

• Access per organisation and user: The mode is enabled for entitled organisations and users. When access is revoked, access and content are removed.
• Legal instruction: Users see and acknowledge the legal instruction (Belehrung) before first use.
• Separate knowledge: §203 content is stored in isolated knowledge storage, with its own chats, agents, and attachments.
• Restricted models: Only sovereign models on approved providers are available. The model picker shows only those models.
• No external access: Web search and connectors are switched off inside §203 mode, and conversation titles never contain prompt text.
• Deletion: Deleting a §203 chat also removes its knowledge content and stored files.

This release corrects how the monthly seat credit is added on yearly subscriptions.

Fixes​

• First credit after trial: The initial seat credit is now added when a trial converts to a yearly subscription.
• Monthly credit: The monthly seat credit is now added from the second month onwards on yearly subscriptions.

Improvements​

• Credit history: Seat allowances now appear as separate entries in your credit history.
• Credit periods: Seat credits follow calendar months based on the start of your billing period.

We tightened browser security for both the chat and the admin application.

Improvements​

• Content Security Policy: Both applications now enforce a Content Security Policy and additional security headers.
• Cookies: Session cookies always carry the Secure flag.
• Agent instructions: Rich-text styling in agent instructions is now removed as part of stricter HTML sanitising.

Zentre now shows a notice that you are interacting with an AI system, as required by Article 50(1) of the EU AI Act. The notice appears on the empty chat screen and under the prompt box, including in agent preview. It is accessible and available in all supported languages.

Fixes​

• Default model: The default team of a new organisation now starts with the intended default model.

Building and sharing agents is now faster and clearer.

Agent preview​

Test an agent while you build it. A live test chat is available when you create an agent, on the agent detail page, and on shared agents. It opens as a side panel on desktop and as a bottom sheet on mobile.

Redesigned Agent Builder​

The Agent Builder has a new layout for model and capabilities, tone, connectors, and instructions. The agent pages have been redesigned to match.

Agent access​

Shared agents open in a read-only view. You can duplicate a shared agent, manage member access in a dedicated panel, and open an info sheet for every agent.

Suggested follow-up questions​

After each response, Zentre suggests three to five follow-up questions that fit the conversation.

Fixes​

• Sharing knowledge files: Sharing several files at once now shares all selected files, not only the first.
• Duplicating shared agents: A duplicated agent no longer inherits a model or knowledge you cannot access.
• Selections: Selected knowledge files and connectors are kept when you switch pages.
• XML upload: Uploading an XML file after "Open Chat" in agent preview no longer fails.

You can now create a Zentre account yourself. The new signup screen is available in all supported languages and confirms by email once your registration is sent.

Fixes​

• Knowledge search: Fixed an intermittent error during knowledge search.
• File questions: Asking "What is in these?" now summarises the file content instead of listing file names.
• Extended Thinking: Claude Sonnet 4.6 no longer returns an error when Extended Thinking is enabled.
• Billing: Fixed several issues with expired plans, trial seats, the checkout preview on reactivation, and checkout currency.

Chat animation​

Starting a conversation feels smoother. When you send your first prompt, the input area moves into the conversation view while the response and the conversation title are generated in parallel.

Typing during a response​

You can keep typing your next prompt while a response is still being generated. You can also attach files and prepare follow-up messages without waiting.

Fixes​

• Teams: User search on the Teams page now works for names with spaces, and the team user count no longer changes with search results.
• Chat: Fixed duplicate responses, repeated knowledge search text, and an input box that stopped responding until the page was refreshed.
• Knowledge: Requests for the document list now return the full list.
• New accounts: The default team now receives all active models, even when some configured models have expired.

Invited users now always belong to a team. The General team is selected by default, and administrators can assign several teams in the same step.

Improvements​

• Team deletion: The confirmation message explains more clearly what happens to team members.
• Knowledge upload: Members without upload permission now see the upload icon with a tooltip that explains why it is unavailable.
• Model selection: The last selected model can no longer be deselected, and an error message appears immediately.

Fixes​

• Organisation preferences: Users promoted from Member to Super Admin can now update organisation preferences.
• Provider names: Model provider names are now shown correctly as Mistral AI and OpenAI.

You can now search across all your conversations in chat.

Improvements​

• Language selector: The login pages for chat and admin now include a language selector.
• Audit search: Searching the audit log has been improved.
• Teams: Teams can no longer be created or edited without an assigned model.

When you paste more than 5,000 characters into chat, Zentre now converts the text into a file and adds it to your knowledge automatically.

Improvements​

• Chat input: The input is focused automatically when you start a new chat or switch between chats.
• Login: The login page language can be set through the link.

Zentre 2.0 is live. 🚀

This release brings a new foundation for the product:

• New interface: An improved user experience across chat and admin.
• EU routing: Requests are routed within the EU.
• Web search: Models can search the web for current information.
• User-based subscriptions: Subscriptions are managed per user.
• Knowledge: Improved stability and quality for knowledge features.