DEVONthink for Mac

This release helps mitigate AI risks, especially when processing online content, to reduce the potential for malicious prompt behavior. We also updated the supported AI providers, adding support for Google's Nano Banana image engine and Anthropic's Claude 4.5 Sonnet.

DEVONthink now detects empty pages when clipping PDFs, ensuring more consistent and useful results, and we made the Safari and share extensions more reliable on macOS Tahoe. Mailboxes in the Import sidebar load faster and match Apple Mail's appearance and you can modify a document's searchable text via the chat or scripting. On the sync side, DEVONthink now automatically adjusts resource usage based on your machine's temperature and workload, and sync in general is more performant. Of course, we fixed bugs and glitches, too.

NEW

• We added independent buttons for the Insert Back Link and Insert Quote commands to the Annotations & Reminders > Annotations inspector.

NEW (Pro)

• We added two new options to the AI > Chat settings: Allow download of web pages and Load remote images automatically to let DEVONthink download and examine web content while generating a response. We disabled these options by default, as they could be considered insecure.
• We added separate Destination settings for image recognition and multimedia transcription.
• DEVONthink now suports Anthropic's Claude 4.5 Sonnet model with pricing and possibilities identical to Claude 4 Sonnet. In addition, the Script Assistant uses now this model instead of Claude 4 Sonnet.
• We added support for Google's Nano Banana text-to-image engine in all image creation functions, e.g., the Generate Image window. This model supports image editing, and, unusually, defines the size and style directly from the prompt. You must have a Gemini API key to access this model.

Improved

• We improved the consistency of syntax colorizing for quotes in Markdown documents.
• When converting EPUB documents, clipping web content, or using the download PDF document from AppleScript command to create PDF documents, DEVONthink now detects empty black and white pages and either filters them out or aborts the clipping.
• When you select a local or global smart group or smart rule in the Navigate sidebar, DEVONthink now uses the targeted group as the destination for new items and grouping selected items.
• We reduced resources used for versioning by 90%, especially during toolbar searches and while updating smart groups and rules.
• DEVONthink now explicitly registers its extensions to ensure they're available in Safari and macOS' Share menu on macOS Tahoe or later.
• You can now modify the plain text AppleScript property of transcribed files, e.g., images, PDF documents, audio and video files. You can also use the transcription synonym.
• The external smart script Zettel-ID handles errors caused by Markdown documents that don't contain this ID.
• We improved the handling of menu scripts execution.

Improved (Pro)

• Apple Mail mailboxes and their structures now load faster and more reliably in the Import > Emails sidebar.
• Cross-linking items is now easier; just drag and drop an item onto the item link pop-up menu in the Info > Data inspector or sheet form views.
• The Help assistant provides the current help page to the chat, enabling summarization, translation, and other AI-assisted tasks.
• The Help and Chat assistants retain unsent prompts when reopening windows, inspectors, or popovers.
• A revised system prompt for chats eliminates unnecessary AI remarks about anonymized links.
• You can optionally save images created or edited via the Chat assistant to the database via prompt. DEVONthink adds the prompt to the Finder Comments of the image.
• The Chat assistant lets you modify the searchable text of transcribed documents (e.g., images, PDFs, audio, and video files), allowing updates to translated or spell-checked versions.
• The context window of the current AI model now determines the number of tags available for the Files > Tags > Import > Use only already existing tags of a database setting.
• Database searches are faster across the board, from the Chat assistant to smart actions and scripting, with notable gains for local AI models.
• Web and Wikipedia searches via the Chat assistant, AI smart actions, and AppleScript commands now also use the Ecosia and Yahoo search engines and better filter results from Bing and Brave.
• All AI responses, whether from summarization, transformation, script commands, or the Chat assistant, now undergo sanitization to remove potentially insecure content. Tool call names are stripped from the input, web page downloads are restricted to non-document queries, and the Chat assistant disables downloads after accessing item contents, annotations, or comments.
• DEVONthink more reliably handles broken links and item links, e.g, returned by OpenAI's gpt-oss model.
• Google's Gemini models now automatically use the latest versions of their Flash Lite, Flash, and Pro models
• The The Devonian Times and Research Overview templates now default to Claude 4.5 Sonnet when Claude is the selected AI provider.
• All Perplexity Sonar models now support vision capabilities.
• The Magistral Small and Medium models use version 1.2 now, adding vision support and expanding the context window to 128,000 tokens.
• Tool calls with Mistral and local AI are simplified and improved.
• Ollama (Remote) now supports API keys, e.g., for the new Ollama Cloud.
• The OpenAI (Compatible) option now works with a broader range of providers and models, including those managed by universities and other institutions.
• Qwen Code Plus is now available via OpenRouter in the Script Assistant.
• The download image with prompt AppleScript command clearly reports an error if the provided image data is not an image.
• We improved the as property descriptions for the chat response for message and add custom meta data commands.

Improved (Sync)

• The Reading List sidebar and custom metadata now sync more reliably via Bonjour.
• DEVONthink now detects the battery and thermal state of the machine and automatically throttles CPU and I/O load during synchronization when necessary.

Fixed

• Thumbnailing many small documents could sometimes cause a crash. Fixed.
• Converting rich text documents or emails to paginated PDF documents could intermittently freeze on macOS Tahoe. Fixed.
• Invalid relative image and transclusion links in Markdown could break document previewing or conversion, or crash the application. Fixed.
• Links entered in lists in Markdown documents didn't work properly in WYSIWYG mode. Fixed.
• Markdown headings containing code didn't display with correct colors. Fixed.
• Links pasted into Markdown lists did not use the correct color. Fixed.
• Searches with proximity operators didn't always highlight hits in PDF documents. Fixed.
• In macOS dark mode, text entered in new HTML documents didn't invert to white. Fixed.
• Tools > Summarize Annotations and the summarize annotations of records AppleScript command sometimes returned annotations in the wrong order, such as those made by PSPDFKit. Fixed.
• Web clipping via the Sorter when opened with the keyboard shortcut incorrectly included Safari's profile name. Fixed.
• Images dropped or pasted into Markdown documents with the Files > Markdown > Import images to group setting enabled didn't apply the Files > Import > Recognition or Files > Tags > Import settings. Fixed.
• Documents moved or duplicated between databases incorrectly added the extension to the internal name and document could be renamed with an empty name if General > Appearance > Show filename extensions was enabled. Fixed.
• Duplicate notifications from the Extras sidebar occurred when multiple main windows were open. Fixed.
• The AppleScript dictionary incorrectly defined and described the root of main windows. Fixed.
• The internal name of files created via AppleScript incorrectly retained the extension. Fixed.

Fixed (Pro)

• Loading the Model pop-up menu in the AI > Chat settings could sometimes cause a crash with local AI providers. Fixed.
• Claude AI wasn't working anymore in the Help assistant. Fixed.
• Retrying a chat in the Help assistant failed after reopening the panel. Fixed.
• The Chat assistant and chat response popovers, such as Summarize, didn't render code blocks or multiple tildes correctly. Fixed.
• The Reveal command in the item link pop-up menus of the Data inspector or sheet form view always opened a new window. Fixed.
• AI responses in the Chat assistant, AI smart actions, or the get chat response for message AppleScript command could stall if required items became unavailable, such as when an item was deleted or its database closed. Fixed.
• Image engines supporting image editing stopped working after creating two images. Fixed.
• The OpenAI (Compatible) option in the AI > Chat > Provider settings incorrectly required /v1/ in the URL. Fixed.
• In the German localization, the OCR > PDF Resolution setting couldn't be changed. Fixed.
• In batch processing, the popovers in the Chat - Query and Chat - Continue if smart actions failed to initialize correctly after switching configurations. Fixed.

Fixed (Sync)

• Bonjour synchronization could sometimes cause memory issues. Fixed.

Removed (Pro)

• We replaced no longer functioning models and outdated models from the OpenRouter options.