DEVONthink for Mac

Here it is, the first feature release for DEVONthink 4. Choose from a much wider range of server-based AI models via OpenRouter or by connecting to just any OpenAI-compatible provider. Integrate models from, e.g., Amazon, Meta, Microsoft, xAI, Baidu, Moonshot AI, Qwen, and more. In addition, DEVONthink Europa adds native support for Claude 4.1 Opus, Gemini 2.5 Flash Lite, OpenAI GPT-5, Google Imagen 4, and the latest models by Mistral. Connect to your institution's Ollama installation with a new configuration option.

To further automate your work, describe your task to the the AI-powered script assistant. It generates AppleScript and JavaScript for Automation (JXA) scripts and even corrects common scripting issues. Edit images, both generated or dropped in from other sources, with prompts using common AI image generation models. As generative AI is notoriously bad at maths, the Chat and other AI-assisted functions now support mathematical functions, and if you're working in academia, retrieval-augmented generation (RAG) gets direct access to the huge document collection on arXiv. In addition, DEVONthink Europa comes with a new Quick Look extension for previewing Spotlight results on macOS Sequoia and later, and more responsive editing of especially larger Markdown documents. It also updates and sorts lists and empties the trash significantly faster.

NOTE: This version is named after Europa, Jupiter's fourth-largest moon. It is a captivating world encased in an icy shell that hides a global subsurface ocean, one of the most promising locations for extraterrestrial life in our solar system. Discovered in 1610 by Galileo Galilei, its name, rooted in Greek mythology, reflects its enduring allure as a symbol of curiosity and the unknown.

NEW

• A new script editor, opened via Data > New > Script lets you draft different DEVONthink scripts, e.g., for reminders or smart rules.
• A new Quick Look plugin lets you preview documents in, e.g., Spotlight searches on macOS Sequoia and later. This only supports standard file formats: rich and plain text, PDF, HTML, XML, images, audio, and video.
• A new search criterion Kind is PDF+Text serves as a shortened form of Kind is PDF/PS and Word Count > 0. Additionally, the toolbar search syntax is kind:pdf+text, its negated form kind:!pdf+text.
• Added a placeholder to the alarm text field in the Info > Annotations & Reminders inspector.
• The Zettel-ID external script for the Script with Input/Output smart rule action pull this information from Markdown metadata.
• The AppleScript record class supports the new read-only properties language and markdown source.

NEW (Pro)

• Support for OpenRouter is available in the AI > Chat settings and the AppleScript suite. This provides access to some AI models that aren't natively supported. Currently accessible models are from: Amazon, Anthropic, Google, Meta, Microsoft, Mistral, OpenAI, Perplexity, xAI, Baidu, Deepseek, Moonshot AI, Qwen, zAI, noting we aren't providing access to beta, preview, experimental, outdated, or distilled models. The Help assistant, as well as the Research Overview and Mixture of Agents - Text/Markdown templates, also support OpenRouter.
• Added support for OpenAI's GPT-5 models. Currently much less expensive than the 4.1 models, it does have a much smaller context window of 400,000 tokens but supports more output tokens and has been trained on more up to date information. Due to the lower costs, web searches are filtered with GTP-5 Nano.
• Added support for Anthropic's Claude 4.1 Opus.
• Added support for Mistral's Voxtral transcription to the AI > Transcription settings. Currently, this uses Voxtral Mini and is less expensive than OpenAI's speech-to-text models.
• Added support for Google's Imagen 4 Ultra in the AI > Image Generation settings, the Generate Image window, and accessible via the download image for prompt AppleScript command. A Replicate.com account is required to use this model.
• Added vision support to the Perplexity models Sonar, Sonar Pro, and Sonar Reasoning.
• Added Ollama (Remote) the Chat > AI > Provider settings and renamed "Ollama" to Ollama (Local). This allows you to access Ollama running on a network, e.g., at a company or university. The remote provider is also supported in AppleScript, e.g. by the get chat response for message command.
• Added a script assistant, available in the Script Editor window, smart actions Apply Script and Script with Input/Output, or reminder scripts. Describe the action you need and a code-compatible AI model will provide a draft script to get things started.
• Save all images in a chat with the Save All Images context menu of the Chat assistant, available if the current chat contains more than one image.
• The Generate Image window can now modify a generated image or one dropped or pasted into the window. Also, images generated via the Chat assistant can be modified. And as the Chat popover is persistent between launches, you can modify an image in it at a later time. Currently, this function is only available with certain models: Flux Schnell, Pro, Pro Ultra, Stable Diffusion Large or Turbo, and GPT-Image-1.
• The Generate Image window now displays a history of created images.
• Given the poor calculations produced by LLMs, a new tool lets the Chat assistamt, smart actions, and AppleScript commands evaluate mathematical expressions.
• Added the new The Devonian Times template, summarizing today's news via AI.
• Added the arXiv site as a search option the the AI > Chat settings.
• Added a Language menu to the AI > Summarization settings.
• Added a new built-in smart rule, Apply Chat Suggestions. This applies seven common AI suggestions to matched documents, e.g., tagging, summarization, etc., utilizing placeholders and one AI request. This rule can be duplicated and modified to suit your needs.
• The Chat assistant, smart actions like Chat - Query, and the download chat response for message AppleScript command can download links in various formats (HTML, Markdown, text, snapshot). The format is automatically chosen, either by the AI or depending on the user's prompt. Currently, this only works with commercial models.
• The new AppleScript property current chat model returns the default model.
• The new get chat capabilities for engine… model AppleScript command retrieves the capabilities of a specific AI model.
• The get chat response for message AppleScript command now has a usage parameter to improve results or control the costs of using external AI.

Improved

• "Audit-proof" databases have been renamed to Revision-proof.
• Editing very long Markdown documents is more performant. DEVONthink detects links in Markdown documents at least 25 times faster when syntax coloring and WYSIWYG is enabled.
• Markdown WikiLinking now supports [[link|-]] to use the title of a linked item. This can be useful when using short aliases for linking.
• Prism.js renders Objective-C code marked with the alternate objc or language-objc as well as the default language-objectivec.
• DEVONthink indexes web-based documents and converts them to plain text better.
• The Go > To Document and Edit > Insert > Item Link popovers now search after a short delay to avoid interfering with typing.
• Updating or sorting by name or kind in item lists is up to ten times faster, including in popovers such as Go > To Document.
• Editing annotations in the Info > Annotations & Reminders inspector now supports versioning.
• The Data > Move > Put Back command as well as its context menu counterpart display the destination when only one item is selected.
• For quicker sorting, the Ascending and Descending commands are now at the top of the View > Sort menu and the sort popup menu in the Information bar of main windows.
• The Add and Preview buttons in the Sorter now display their shortcut keys when you hold the Command key.
• In the Sorter's Voice Note and Video Note tabs, the camera and microphone menu is disabled if DEVONthink is disabled in macOS' Privacy & Security > Camera settings.
• Emptying the Trash is up to five times faster when versioning is not enabled.
• DEVONthink's built-in folder action scripts now run smart rules using the perform smart rule AppleScript command with the On Import and On OCR event triggers, as appropriate.
• Querying tags via AppleScript is now supported, e.g., tags contains "support" or where its tags contains {"2025","documentation"}.
• Records now support the sheet-related AppleScript commands get cell at row/column, set cell at row/column, add row cells, and delete row at position.
• The AppleScript database property "audit proof" is now revision proof. However, audit proof has been added as a synonym to ensure backwards-compatibility.

Improved (Pro)

• Summarizing documents to annotations, e.g., via Edit > Summarize via Chat or Tools > Summarize Documents via Chat, first saves a version when versioning is enabled.
• The Research Overview template now supports local AI engines and models having tool call support and a context window of at least 32,000 tokens.
• The notification of a chat-summarized RSS article is simpler when only one article is involved.
• The search assistant now accepts local AI models with a context window of at least 8,000 tokens.
• In the Import > Image Capture sidebar, the Group menu now has a search field for quickly locating groups. Also, the destination settings are saved after each scan and restored when reconnecting to the scanner.
• The From date in the Import > Emails sidebar now defaults to the last used To date.
• The Chat assistant better handles detecting selected documents, especially when using the Chat popover.
• AI-directed features using web, Wikipedia, and PubMed searches are more efficient, filtering out excess or links already processed by concurrent searches. This affects e.g., the Chat assistant, smart actions, and AppleScript commands, and could speed up responses and minimize token usage.
• The Chat assistant's progress indicator more clearly shows chat activity, e.g., searching or downloading resources.
• The Chat assistant doesn't scroll a received chat response out of sight anymore.
• Prompt fields across the application, e.g., in the Chat assistant, the Generate Image window, and the Help assistant, now support using ⇧↩ in addition to ⌥↩ for inserting a line break.
• The model menu in the Chat assistant features a search field if more than 25 items are available.
• Search request made to Google, Bing, and Brave via chat functions, e.g., the Chat - Query action or prompts in the Chat assistant are more reliable. DEVONthink also handles Mistral's rate limits better.
• Bulk processing items via chat, e.g., tagging or summarizing, uses better error handling.
• Chat tools calls for local AI models have a reduced temperature for more reliable results and can now download web pages as needed. In addition, the Chat assistant handles errors better.
• Irrelevant PubMed and arXiv search results are now filtered in all AI functions, e.g., in the Chat - Query smart action, maintaining the context of the current conversation. For commercial providers, the least expensive model is used for filtering. Additionally, once enough results have been received from these site and Wikipedia, other pending evaluations are cancelled, saving both tokens and time.
• Chat suggestions in smart actions, e.g., suggested tags, are more reliably retrieved.
• The download image for prompt AppleScript command now reports an error if the specified model doesn't support image editing. When using the Flux engine, the Flux Kontext models are used as they are built for image editing, not just generation.
• Updated Gemini 2.5 Flash to the final version.
• Both filtering web search results and the Help assistant use now Gemini 2.5 Flash Lite.
• Updated Google Image 4 to final version and is used in the Mixture-of-Agents - Images template.
• Both AI > Mixture-of-Agents templates no longer use Perplexity due to an increase in irrelevant responses.
• Updated Mistral Medium to version 3.1. Mistral Large 3 has better vision support. And "Mistral AI" has been renamed to Mistral in the AI settings and Chat assistant.
• The default context window set by DEVONthink for Ollama is now 16,000 tokens.
• The "GPT4All" provider has been renamed to OpenAI (Compatible) and can be used connect to any compatible remote services. It also now supports an API key, affecting the AI settings, menus and related AppleScript commands.
• Improved compatibility of chat-based smart actions with LM Studio's MLX models, e.g., Gemman 3n.

Improved (Server)

• Improved HTTP handling, should also address a rare crash.

Fixed

• Markdown lists didn't render properly when indented with one space. Fixed.
• Thumbnailing Markdown documents could cause a rare crash. Fixed.
• Scrolling Markdown documents was affected after pasting copied text, changing margins, or syntax highlighting. Fixed.
• Remote images in Markdown documents didn't render in WYSIWYG mode if the files didn't have a valid file extension. Fixed.
• Markdown MathJax equations in Asian text didn't render properly. Fixed.
• Indexed Google Drive documents couldn't be opened via the Data > Open With command. Fixed.
• Links using Apple Notes' URL scheme couldn't be clicked in documents. Fixed.
• For RSS and Atom feeds with multiple authors, only the first was used. Fixed.
• An open Quick Look panel didn't update when changing the selection of the view/edit pane when General > Interface > Preview group content was enabled. Fixed.
• In the item list in List view, the Tags column didn't display autocompletion results. Fixed.
• The tooltip shown when hovering over items with truncated names didn't fully display the name if the item also had property icons. Fixed.
• Groups with a custom icon and applied color displayed differently in different parts of the interface, e.g., the item list and context menu. Fixed.
• In the Navigate sidebar, you could only drag and drop multiple favorites, smart groups, or smart rules by the icon, not the name. Fixed.
• The First, Previous, Next, and Last commands in the Go menu couldn't be used in toolbar search results. Fixed.
• The search field in the context menu and destination popups, e.g., in the Sorter, didn't sort the results by relevance. Fixed.
• In a multi-monitor setup with vertically positioned screens, the Sorter was only partially visible as a docked tab. Fixed.
• The item count in the Database Properties window incorrectly included saved versions. Fixed.
• The internal versions group was incorrectly made visible in the Sorter. Fixed.
• The Spotlight index for a database could incorrectly contain document versions. Fixed, though rebuilding the Spotlight index via Database Properties may be needed.
• .strings files weren't indexed correctly, potentially affecting versioning of AppleScript modified files. Fixed.
• The Scan Text smart rule action didn't ignore Markdown metadata if there was no other content in the document. Fixed.
• If the DEVONthink helper background task crashed, other background tasks, e.g., running scripts, could be blocked for a period of time. Fixed.

Fixed (Pro)

• Processing multiple documents via the Chat assistant could sometimes crash the app. Fixed.
• Timestamps weren't properly restored when revisiting the Chat assistant. Fixed.
• Replies in the Chat assistant could contain improperly escaped Markdown. Fixed.
• The (AI) suffix on an AI-generated document names could be misplaced if the name already had a file extension. Fixed.
• Claude 3 Haiku responses in the Help assistant often included invalid links. Fixed.
• Switching to Mistral in the Help assistant and resending the last chat message failed. Fixed.
• AI-generated web searches containing multiple site: operators didn't work correctly. Fixed.
• Item links or WikiLinks to groups weren't supported in the AI > Graph inspector. Fixed.
• The AI > Graph inspector could incorrectly display a connection line to a hidden (disabled) node. Fixed.
• In the popup of URL fields in the Info > Data inspector and the form view of sheets, the Paste and Remove commands neither saved changes nor updated the layout. Fixed.
• In the Import > Emails sidebar, nested iCloud mailboxes were duplicated at the account's top-level. Fixed.
• Search queries using date-based custom metadata weren't always parsed correctly. Fixed.
• Some item properties, e.g., the URL or UUID, couldn't be retrieved by AI in the Chat assistant, smart actions, or the get chat response for AppleScript command. Fixed.
• The prompt for Scripts > Chat > Geolocate Image was incorrect. Fixed.
• The get chat response for message… as HTML AppleScript command could return incorrectly mapped links. Fixed.
• Gemini 2.5 Pro didn't work with the get chat response for prompt AppleScript command when thinking was disabled. Fixed.

Fixed (Server)

• The icons in the Server > Users settings didn't look correctly in dark mode. Fixed.

Removed (Pro)

• Removed Data > New from Template > AI > Script Assistant as it has been replaced by the Script Editor and Assistant.
• Removed support for Google's Imagen 3 (Fast) AI model.
• Removed support for Anthropic's Claude 3.7 Sonnet and Claude 4 Opus AI models.