What we shipped recently

Universal search (⌘K)

One search for the whole platform: chats, assistants, workflows, prompts, apps, projects and settings now live in a single modal — open it with ⌘K or the new Search button in the sidebar. Arrow keys drill straight into a category; selecting a hit opens or runs it instantly.

Redesigned overview pages

The chat overview has been rebuilt from the ground up — with search, sorting by latest activity, and a calmer, more consistent page layout. It is the first page on the new universal page framework, which will roll out to more areas over time.

Tool calling in the OpenAI-compatible API

The OpenAI-compatible API now supports tool calling: pass tools and tool_choice and get tool_calls back — streaming or not. Agents like OpenCode work with meinGPT out of the box. Models without tool support (e.g. Perplexity) now return a clear error instead of silently ignoring the tools.

Fixes & Improvements

• Knowledge & skills in assistants now save reliably — uploaded files such as PDF or Excel stay in place after you save and reopen the assistant.
• CSV files in UTF-16 encoding are now detected and read correctly on upload instead of arriving as garbled characters.
• Steadier chat responses when tools run in parallel — short-lived database conflicts are retried automatically instead of surfacing as errors.
• Further stability and performance improvements in DataVault — concurrent pool syncs no longer block one another, embedding services respect rate-limit cooldowns, and large files process more reliably.

Italian as a new language

meinGPT is now fully available in Italian — the entire interface, all Academy courses, use cases, and the Outlook add-in render automatically in Italian whenever the profile language is set to Italiano.

Fixes & Improvements

• More reliable meeting bot — the recording bot now detects on its own when a machine gets stuck, replaces it automatically, and works through any backlog. Meetings no longer drop because a recording machine ran full or got blocked.
• Steadier DataVault search — keyword search now retries rare, transient database errors automatically instead of failing the result.
• Web pages with very long titles now index reliably during crawling instead of failing with an error.
• Further stability and performance improvements in DataVault — Knowledge Insights refresh faster in the background, and services starting concurrently no longer trigger lock-timeout restarts.

Shared slide templates

Admins can now promote slide templates into a shared library for the entire workspace. In settings, templates can be moved between "Shared" and "My templates". All workspace members see shared templates read-only and can use them in assistants and workflows.

French as a new language

meinGPT is now fully available in French — the entire interface, every Academy course, use case, and the Outlook add-in appear automatically in French when the profile is set to Français.

Workshop materials as files

Workshop materials can now be uploaded as PDF or DOCX, not just pasted as links. Files are stored securely and downloadable by participants straight from the invitation or follow-up email — even without a meinGPT account.

OpenAI-compatible /v1/models API

Any OpenAI-compatible SDK can now query /v1/models against meinGPT to list available models — useful for custom integrations and automations.

For administrators

• Shared MCP connections now appear in personal connection settings, including the assistants that use them. The connection can be disconnected there but not edited.
• SSO enforcement on signup — new email accounts can no longer be created on domains where an organization requires SSO. Affected users are directed to the SSO login instead.
• Clean up orphaned data pools — pools already removed in meinGPT can now still be cleaned up in Vault to delete leftover data.

Fixes & Improvements

• GitHub connector fixed — the connector can now be connected via OAuth. Once authorized, the assistant calls the GitHub tools for repositories, issues, and pull requests instead of answering from memory.
• Tool-using chats are more reliable — long, multi-step tool conversations no longer fail with an occasional error partway through.
• Calculator, memory, and Google Maps tools now work reliably in assistants.
• Pasting text with asterisks, underscores, or tildes (e.g. code or formulas) now preserves those characters.
• Shared chats now show images generated in the chat to everyone the chat is shared with.
• Workshop improvements: calendar invites and cancellations now work for external recipients (e.g. Gmail), trainers are notified by email when assigned and shown in the Learn hub, multiple participants can be invited at once.
• DataVault processes large files from staging storage more reliably.

Four new models

Four new models are available — administrators can enable them in the organization settings:

• Claude Opus 4.8 — Anthropic's latest Opus with a 1M token context, sharper agentic coding, and stronger reasoning. EU-hosted via Vertex AI.
• Gemini 3.5 Flash — combines pro-level reasoning with high speed and understands text, image, audio, and video input. Enabled automatically for organizations already running Gemini-family models.
• Mistral Medium 3.5 — latest generation of the EU-hosted Mistral Medium model.
• Kimi K2.6 — successor to Kimi K2.5, hosted on Azure AI Foundry (Sweden).

Pick connectors directly in chat

Connectors like Google Calendar, Microsoft Teams, or custom MCP servers can now be selected directly from the @ menu in the chat input. Each pick activates the connector and inserts an @-mention in the message — same pattern as assistants and data pools. Active connectors appear as chips below the input and can be removed there with one click.

Import memories from other AIs

Settings → Memory has a new import wizard. Paste a prompt into ChatGPT, Gemini, or Claude, paste the answer back into meinGPT, review the extracted memories, and apply with one click. Carrying work context over from another AI now takes a few seconds.

Fixes & Improvements

• Per-function permissions for connectors like Outlook or Teams are now reliably honored when the connector is activated directly from a chat.
• Steadier chat stream on brief disconnects — When the tab switches or the network blips, the response now picks up live immediately on reconnect instead of staying silent until the next larger step.
• No more duplicate messages when the document or canvas view opens during an in-flight response.
• Excel Agent entry in the sidebar — Icon is properly centered in the collapsed sidebar; the hint text now shows up directly below the row instead of far to the right.
• Very large documents stay searchable — Documents that exceed the chunk cap are now automatically indexed via keyword search (BM25) instead of failing. They remain searchable, just without semantic similarity matching.
• Documents with tokenizer control strings index reliably — Documents that contain strings like <|endoftext|> as normal text (e.g. API docs, AI transcripts) no longer fail during token counting.

New spreadsheet editing in chat

Spreadsheets generated in a chat now open directly in the right-side drawer for editing — no download, no Excel needed. The code sandbox ships with a modernized xlsx skill that uses the preinstalled libraries (pandas, openpyxl) consistently and produces more predictable results. A new partial-read tool lets the AI fetch targeted slices of large files instead of loading the whole thing. The legacy Excel agent has been retired as part of this move — the code sandbox is the new default everywhere.

Unified chat input

Sources, tools, and tables now live in one shared dropdown below the input field instead of several side by side. A per-chat tables setting in that same dropdown controls whether the AI may use table formats.

New sidebar design

The sidebar has been rebuilt on shadcn primitives: more consistent spacing, steadier collapsing and expanding, and a tidier look in both states.

For administrators

• Credit top-up by invoice works again — Invoices for credit top-ups are issued correctly again; VAT is transmitted in the expected format.
• Correct month boundaries in the usage export — The monthly sheet calculates in a consistent timezone regardless of DST switches.

Fixes & Improvements

• The "Execute" tool no longer truncates long output.
• The thinking bubble now appears while the chat waits for the first stream chunk, not only when the first character arrives.
• Datapool sync is more resilient to single-file failures (crawler, chunker, retry behavior).
• Sandbox code panels in the chat history are no longer re-rendered on every new message.
• Tidier tool steps in chat — Tool calls render more consistently, with a clearer layout and a unified file-icon set.
• Slides templates in the account section — The Slides templates page now lives under account settings and uses informal phrasing throughout.
• Send button hidden in workflow steps — Workflow steps without a chat input no longer show a send button or dictation button.

What's new

• Enable connectors directly from chat — Connections to services like Gmail, Google Calendar, Asana, or Confluence are now set up right from the chat the moment the assistant needs them. No detour through settings to finish a task.
• A clearer plan page — The subscription page now shows, per cost class, what is included in the plan and what is billed by usage. A new "Request a quote" button reaches the sales team with full workspace context.
• Usage export, redesigned — The export now consists of four combinable Excel sheets: per-user chargeback, a monthly summary for invoice reconciliation, a User × Model matrix in euros, and a bookkeeping sheet for top-ups and expirations.
• Native image vision in the code sandbox — Screenshots, charts, and individual PDF pages in the code sandbox are now read natively instead of being guessed via OCR. Multiple files transfer to the sandbox in a single step.

Chat

• Tool steps in the conversation language — The short descriptions above each tool step now appear in the language of the conversation. Chatting in German now shows the intermediate steps in German instead of English.
• Sent messages stay in place — A message you send appears directly at its final position instead of briefly showing up at the bottom of the chat and sliding upward before the response starts streaming.

Code sandbox

• The most-used packages are preinstalled — Data analysis, PDF processing, and image tasks start immediately. Preinstalled packages include pandas, numpy, matplotlib, Pillow, reportlab, several PDF and OCR libraries, plus SheetJS and pdfkit for Node. Repeated pip install steps at session start are gone.

For administrators

• Billing is visible per cost class — Standard, Premium, and API usage each carry their own status (included, billed by usage, or blocked). Sales and support see at a glance what is metered and what is bundled into the plan.
• Four mix-and-match views in the usage export — Per-user chargeback (default), an invoice-reconcilable monthly summary with opening and closing balance, a User × Model matrix in euros, and a bookkeeping sheet for top-ups and expirations. Multiple views fit into a single Excel file.
• "Request a quote" directly from the plan page — The button sends a structured message to the sales team, including workspace context. The previous detour through a Fillout form is gone.

Fixes

• Mandatory-course modal no longer covers content — Opening a mandatory course used to be hidden by the outstanding-courses modal. The modal is now hidden on course pages, and the button jumps directly to the chosen course.

What's new

• Auto top-ups for credits — Workspaces with a saved card can now set up automatic top-ups: when the balance drops below your threshold, meinGPT charges your saved card automatically. Available under Settings → Subscription.
• Buy credits on invoice — Workspaces on invoice billing now buy credit packages (€500–€10000) directly from the subscription page. Credits activate immediately, the invoice arrives by email from our accounting team.
• Usage and top-up history with export — The subscription settings now show a monthly overview and a unified payment and top-up history. Clicking a month filters the daily chart, each row offers a direct export. Usage can also be exported as Excel or CSV for a selected date range.
• Slides in assistants and workflows — Slide templates can now be attached to assistants and workflow steps. Slides created from chat or from automations use your organization's selected templates end-to-end.

Chat

• Tool steps collapse into one card — Several consecutive tool calls in one reply are now grouped into a single collapsible card. While running you follow the progress live; once finished the card folds back with a total duration — less scrolling, clearer overview.
• Thinking time and tool durations are visible — A new indicator at the top of a reply shows how long the assistant is thinking; each tool call shows its run time inline. The final "Thought for X seconds" line stays visible even after a page reload.
• Sonnet 3.7 replaced by Haiku 4.5 — Claude Sonnet 3.7 (incl. the Thinking variant) has been retired by Anthropic. Existing assistants and workflows using it are switched to the newer and cheaper Claude Haiku 4.5.

Learning & onboarding

• Academy in English and Spanish — All 80 Academy courses are now fully available in English and Spanish. Course language follows the profile setting; users on other locales fall back to English.

Data & knowledge

• PowerPoint import is more reliable — Large slide templates load completely (deeply nested structures previously crashed the import), logos and graphics are preserved, text boxes no longer show a gray background, and shapes no longer pick up black borders that were not in the original file.
• OneDrive: no needless full scans on empty sources — Empty OneDrive data sources no longer start periodic full scans. Sync stays quieter and uses less background work.
• OneDrive: auto-repair of older file entries — File records without a technical ID are detected and repaired on the next sync. Future OneDrive deletions remove the files from the Vault index more reliably.
• Deleted files disappear more reliably from the index — During brief database issues the cleanup is now retried instead of leaving stale search results behind — applies to files from every connector and to deleted pages from web-crawler sources.

For administrators

• More reliable SSO across all hostnames — Microsoft and Google sign-in no longer shows other customers' URLs in consent screens. White-label workspaces now use a central callback, which makes sign-in via custom hosts more stable.
• Safer workspace domain management — Workspace admins now add their company domain themselves and confirm ownership via a DNS TXT record. The platform suggests likely domains when several members share an email domain.
• Pre-filtered login links — Workspace admins now find a copyable login link in workspace settings that pre-filters the sign-in page to specific methods (e.g. Microsoft SSO only). Useful for teams that should only use one sign-in variant.
• Image-model management for admins — A refreshed settings page groups all available image models per provider, shows quality and pricing, and lets admins enable, disable, or default individual models. The UI for selecting a model in the image generator app has also been revises.
• Perplexity models honor the admin setting — When Perplexity is disabled in workspace settings, Perplexity models (incl. Deep Research) can no longer be picked as the preferred model and no longer return results.
• Manage auto top-ups — A new "Auto top-up" section on the subscription page lets admins pick the card and the threshold. Cards are added through the Stripe Customer Portal.
• Monthly and daily view in the AI Cockpit — Click a month to filter the daily chart; the per-row export delivers Excel or CSV for that month. Manual top-ups and corrections show as their own bucket in the unified payment and top-up history.
• Docs link in slide template settings — On the workspace settings page for slide templates, the subtitle now links directly to the docs with best practices, Mustache placeholders, and the layout naming convention.
• Local Vault installations without Postgres extensions — On-premise Vaults can now run with SQLite and LanceDB locally. Cloud Vaults stay unchanged.

Discontinued

• "Ask IT" is removed — The separate Tech Support app is gone. The same knowledge collections, instructions, and mailto fallback live on in a pinned Assistant — the migration wizard prepares the conversion.

What's new

• @-commands in chat — Typing @ in the chat input opens a quick picker for an assistant, data source, prompt, or workflow. The selection applies only to the current message — persistent chat settings stay unchanged.
• Meeting bot with calendar automation — Google and Microsoft calendars can be connected once. Per user, the bot can be set to join all meetings, externals only, or only meetings with #rec in the title. Existing meetings appear in a picker, new ones are picked up automatically.
• Workshop management in meinGPT — Administrators create workshops and events directly in the workspace, invite participants, and manage capacity. Trainers have a dedicated profile with photo and bio that is maintained once and reused for every workshop — the photo automatically appears in invitation emails.
• More reliable message actions in chat — Copy, regenerate, and feedback buttons under chat replies are now visible on touch devices and in browsers where they were previously hidden under certain conditions. A new chat setting keeps the action bar permanently visible.

Data & knowledge

• SQL Connector works with PostgreSQL and MSSQL again — "Test connection" against Postgres hosts (e.g. Neon) and MSSQL servers no longer fails with an internal error. Existing connectors are unaffected.
• Custom MCP tools with required fields work again — Tool calls with required parameters (e.g. Gamma deck generators or other custom MCPs) now receive the correct arguments and are no longer sent empty to the server.
• OneDrive sync no longer deletes mistakenly "missing" files — For OneDrive connectors with a sub-folder as base path, the delta sync occasionally marked files outside that path as deleted. That no longer happens.
• More stable document processing — Several fixes to the document engine: attachments with identical content no longer block processing, transient Vault outages produce clean retries instead of stuck files, and a single failing document no longer holds up the rest.

Meetings

• Correct speaker names in summaries — Instead of generic labels like "Speaker 1", meeting summaries show the real participant names when they are known from the transcript.

For administrators

• AI Cockpit shows real usage only — Manual credit recharges and adjustments no longer appear in cost charts. Daily, per-user, and per-team views reflect actual AI usage only.
• Workspace switcher filters by active subscription phase — The workspace switcher only shows organizations with an active or upcoming subscription phase. Expired or cancelled workspaces are hidden but remain reachable via workspace settings.
• Workflows hub shows the "+" button consistently — Every project now shows the create-workflow button. For reader roles it appears disabled, with a tooltip explaining the reason.
• Clearer Workflow API examples — The external Workflow API docs show the Authorization header consistently and explain file uploads (including multi-file) more clearly.

Fixes

• Workflow editor shows the correct tab title — The first tab is now reliably labeled "General" instead of showing a misleading label from another area.
• Adoption dashboard no longer breaks on missing workflow data — The workflow matrix stays consistent even without workflow data — empty cells used to shift entire rows.

Bug fixes

• Excel assistant picks up assistant-attached spreadsheets — The Excel assistant now also finds spreadsheets attached to an assistant, not just files freshly uploaded in the chat. Larger analyses also have more time to complete before timing out.
• Meeting transcription: video uploads reliable again — Recordings can be uploaded without aborting again; recently affected uploads were failing before transcription could start.
• Interrupted replies stay visible — If a reply breaks off unexpectedly mid-sentence, you now see what was already written — with an error notice and Retry / Switch Model buttons directly below it.
• Translation tool: small fixes — Norwegian is selectable across providers again. Translated documents get the language suffix back (e.g. report_de.pdf) and download automatically once they are ready.
• Fewer misleading error messages — Expected background failures are less likely to trigger red alerts, while important errors remain visible.
• Web crawler login more reliable — Wrong credentials are now reliably detected (previously some were silently accepted), and Symfony / Pimcore sites with CSRF protection can be connected. Crawled pages get URL-derived filenames, making them easier to trace back to their source.
• Web crawler finds every page reachable from navigation menus — With "only main content" filtering, up to 80% of nav-linked pages were previously invisible to the crawler — they are now fully captured.

What's new

• Web crawler can now crawl password-protected intranets — For each web-crawler source, you can store a login URL, username, and password once. This allows you to connect password-protected internal pages as sources.

Improvements

• More reliable chat answers — Brief hiccups from the model provider are now retried automatically in the background instead of surfacing as errors in the chat.
• More reliable GPT-5 responses — The assistant no longer truncates mid-sentence in chats with email or calendar tools enabled.
• Freshly uploaded files attach more reliably — The chat now briefly waits for text and image processing to finish before the file is handed to the response.

Improvements

• Generated images render reliably in chat — Images now appear directly in the chat as a dedicated card, with buttons to copy or download. While the image is being created you see a clear "Generating image" indicator.
• Clear notice when an answer is interrupted — If the connection drops while a response is being generated, you now see a clear notice in the chat and can retry with one click — instead of the message spinning forever.
• Sidebar polish — The active dot is centred again and hover states between tabs are clean.
• SharePoint and OneDrive sync reliably again — Fixed an automatic token-renewal failure that affected organizations with an internal tenant slug.

For administrators

• Ask IT migration wizard — A new admin wizard turns your existing "Ask IT" setup into a workspace assistant — including knowledge sources and an automatic email draft to your IT support when the assistant can't help. After migration, Ask IT disappears from the sidebar.