DEVONthink for Mac

This maintenance and feature update to DEVONthink brings significant enhancements to Markdown and document editing, web clipping, AI workflows, and overall system performance. Create and edit tables directly in the Markdown WYSIWYG editor without touching raw syntax, take advantage of automatic renumbering for numbered lists, and export documents using a completely rewritten Word engine that faithfully preserves styles, citations, and MathJax formulas. Web clipping and the clutter-free reader view now leverage publisher-provided Markdown streams, retaining article subtitles, image captions, and rich metadata across capture formats.

For Pro and Server editions, DEVONthink 4.4.1 accelerates Apple Mail integration with up to 30 times faster message listing and background hierarchy loading. The AI toolset expands with the new Chronology skill, support for OpenAI's GPT 6 Astra, Sol, and Luna, and GPT-Image-2.5, Claude Fable 5.1 and Opus and Sonnet 5.5, remote Gemini transcription, and search integration with Tavily and Brave. Server editions now allow downloading complete groups from the web interface, powered by an asynchronous engine with virtually no memory overhead for multi-gigabyte transfers. Comprehensive compatibility updates for macOS 26 Tahoe and macOS 27 Golden Gate, along with overall reliability and performance improvements, round out this release.

NEW

• The new Edit Table button in the Editing bar, the context menu, and the Tools > Markdown menu lets you visually create a WYSIWYG Markdown table with a grid picker. Commands and buttons in the same places let you add, move, or remove columns and rows and change column alignment. As in sheets, ↩ in the last row or ⇥ in the last cell adds a new row, and ⇥ and ⇧⇥ move forward and backward between cells.
• The new Edit List button in the Editing bar, the context menu, and the Tools > Markdown menu creates a Markdown list at the cursor, converts selected text to a list, or changes the type of an existing list.
• Ordered lists in Markdown renumber automatically when you insert or remove items, and nested lists update independently of their parent lists.

NEW (Pro)

• A new Chronology skill in the AI > Library settings turns selected documents or the contents of a group into a dated timeline, with each event linked to its source document. It lists contradictions between the sources and keeps undated information separately.
• DEVONthink supports Tavily and Brave Search as providers in the AI > Search settings.
• The AI > Transcription settings offer remote transcription via Google Gemini.

NEW (Server)

• The web interface supports downloading complete groups.

Improved

• File > Export > as Word 2007 document (DOCX) uses a new Markdown renderer that keeps tables, quotes, callouts, images, links, lists, footnotes, and headings as Word styles, links citations to their references, and turns a table of contents into an updatable Word field. It also converts MathJax formulas.
• Web clipping, clutter-free clipping, and the Reader view use a page's published Markdown version when available, which yields complete pages even when a page is rendered with JavaScript.
• The up and down arrow keys move the cursor row by row within the same column of a Markdown WYSIWYG or rich text table.
• Markdown themes define their own color for wiki links, separate from the color of regular links.
• Markdown documents containing many images open and scroll more smoothly in WYSIWYG view. Also, images that finish loading above the current scroll position no longer push the text you are reading out of view.
• Custom Markdown style sheets can override the image scaling of the built-in style sheet.
• Markdown documents written in right-to-left scripts, such as Arabic or Hebrew, open with their writing direction determined by their text instead of the interface language.
• CriticMarkup text appears in the theme's insertion color instead of as plain underlined text in previews and HTML conversions.
• Syntax colorizing in Markdown colors strikethrough text, CriticMarkup deletions, and CriticMarkup insertions the same way as the rendered preview.
• DEVONthink Markdown retains formatting, inline code, and links in sub- and superscript text.
• Citations in rendered Markdown show their number from the references list, e.g., (1) or (p. 42, 1), instead of the cite key and no longer appear in italics. Several citations in one bracket also share a single pair of parentheses.
• DEVONthink Markdown accepts MultiMarkdown's doubled delimiters for mathematical content, \\(…\\) and \\[…\\], so formulas appear typeset instead of as LaTeX source code, including in the Chat assistant.
• Converting rich text to Markdown is more accurate.
• DEVONthink embeds images on clipped web pages that use <picture> or srcset instead of linking to them, so captures no longer depend on the original site.
• Markdown clippings from documentation sites, e.g., Swift-DocC or Sphinx, keep headings as plain headings with their anchor ID instead of turning them into links to their own anchor. Genuine links in headings remain intact.
• Clutter-free clipping adds metadata to PDF, RTF, formatted note, HTML, web archive, and plain-text documents, as it already did for Markdown via frontmatter.
• Clutter-free clipping and the Reader view include article subtitles published by Substack, The Atlantic, The Guardian, and other sites.
• Clipping web pages as Markdown, clutter-free documents, or PDF is faster and more reliable, as DEVONthink captures pages directly instead of using the background download service.
• Clutter-free clipped forum and helpdesk threads start with the thread's title.
• Clutter-free clippings omit empty paragraphs, figures, and lists left behind by cleanup. Markdown conversion also omits empty width and height attributes for lazy-loaded images.
• The Reader view and clutter-free captures omit MediaWiki's "Jump to navigation" and "Jump to search" shortcuts.
• Markdown clippings use a picture's own caption, such as a photo credit, instead of a title that repeats the image description. Metadata for these clippings includes the date when available.
• Items in the trash no longer appear in bold (for example unread or duplicated items), and the strikethrough is drawn lighter, so their names are easier to read.
• Windows report the document's item link through Accessibility, e.g., for screen readers.
• In the Help viewer, the back and forward buttons appear in the content pane, similar to Apple's help, and the toolbar shows the Help Assistant's chat buttons.
• Removing smart rules via the Support Assistant is more reliable.
• In narrow windows, an overflow button at the end of the Navigation and Editing bars shows the tools that don't fit.
• Inspectors have a more clearly defined boundary, with a revised Liquid Glass appearance on macOS 26 Tahoe and a separator line on macOS 27 Golden Gate.
• We modernized several aspects of the interface to follow the current operating system's styling.
• If you launch DEVONthink from its disk image, moving it to the Applications folder ejects the disk image automatically.
• The open window for AppleScript command accepts the URL and referrer parameters, similar to open tab for, and its record parameter is optional. Both commands also accept an HTML source to view, e.g., a dashboard created by AI, with a specified URL serving as the base URL of the source.
• The resolve book metadata AppleScript command and the Pro edition's MCP tool use the latest Open Library API.
• We updated the SkimNotes framework to version 1.6.3.
• Improved overall performance and reliability.

Improved (Pro)

• DEVONthink loads Apple Mail mailboxes and accounts in the Import > Email sidebar 10 to 15 times faster and retrieves the complete structure, including unread counts, in the background instead of level by level. Messages load up to 30 times faster, so large mailboxes with thousands of emails load in minutes. The Priority filter is disabled while an Apple Mail mailbox is selected, as Apple Mail doesn't provide this information.
• The Download Manager has a modernized appearance.
• Opening the Chat inspector, or a window with the inspector visible, automatically loads the most recent chat for the current document or selection.
• The AI > Chat settings validate the API URL as you type, explain errors below the field, and don't save incomplete addresses.
• Changing the server URL of OpenAI (Compatible) providers refreshes the model capabilities instead of reusing results from the previous server. Capability checks are also more robust and ignore temporary failures.
• Leaving the Keyword field of a skill unchanged in the AI > Library settings no longer saves and re-syncs the skill.
• Checking for installed command-line clients, e.g., Claude Code, in the AI > MCP settings stops after a few seconds instead of waiting indefinitely.
• DEVONthink creates its folder in iCloud Drive only when you work with a skill script, e.g., saving or importing one, and doesn't recreate the folder on launch if you delete it.
• The Chat - Query and Chat - Continue if smart actions skip the subsequent actions when an error occurs.
• The extract_record_content MCP tool resumes where the previous call stopped when extracting PDF documents that exceed the time limit, so very long documents can complete over multiple calls.
• The transcribe record AppleScript command and transcribe_record MCP tool support PDF documents with an unusable text layer.
• Chats using tools reuse Anthropic's prompt cache over the whole sequence of tool calls, reducing the cost of long conversations.
• Chats with Claude Sonnet, Opus, and Fable show the model's own progress notes between tool steps, e.g., what it just found and what it intends to do next. Fable also requests independent items in one turn instead of one per turn, e.g., when creating multiple documents, making long conversations using tools faster and cheaper.
• We updated Claude Sonnet and Opus to version 5.5 and Fable to version 5.1, via Anthropic and OpenRouter.
• We updated GPT 5.6 Sol and Luna to version 6, via OpenAI and OpenRouter. Both are more affordable than their predecessors.
• GPT 6 Astra replaces OpenAI's GPT 5.5 (Pro), with a context window of 1,050,000 tokens and a knowledge cutoff of April 2026.
• We updated Gemini Flash to version 3.8, via Google and OpenRouter.
• Image generation and editing via OpenAI use GPT-Image-2.5 at the price of GPT-Image-2, with faster generation and more precise adherence to editing instructions. The AI > Mixture of Agents - Images template can also use this model.
• The script assistant supports the latest models, including Muse Spark 1.3 via OpenRouter.
• Reasoning models in Ollama (0.34.3 or later) think at the depth reported for the model. The Usage popup in the AI settings controls the depth, where Auto uses medium reasoning and Best the deepest level. Vision-capable models in Ollama can also receive images from tool call results, e.g., screen captures, so they can use MCP tools and skills if their context window is at least 65,536 tokens. Currently, Gemma 4, Qwen 3.x, and Muse Glimmer are supported.
• LM Studio (0.4.23 or later) supports vision tool call results, e.g., screen captures. Currently, Gemma 4, Qwen 3.x, and Muse Glimmer are supported, and these models can also use MCP tools and skills if their context window is at least 65,536 tokens.

Improved (Server)

• The web server processes requests asynchronously, so slow tasks, e.g., downloading huge files, no longer stall the application or block other connections.
• Downloading documents and groups from the web interface requires only minimal resources.

Improved (Sync)

• Synchronization no longer reindexes, rewrites, or reloads documents whose received contents are identical to the local copies.

Fixed

• Scrolling could crash DEVONthink on macOS 26 Tahoe and later, especially when scrolling quickly beyond the end of the content. Fixed.
• Editing a document during synchronization could lose text entered after synchronization began. Fixed.
• Reloading a Markdown, HTML, or web archive document could restore the text cursor and scroll position to their locations from the last save. Fixed.
• Embedded YouTube videos in Markdown and HTML-based documents wouldn't play, as YouTube requires embedded players to identify the app via the HTTP referrer. Fixed.
• Editing a syntax-colorized Markdown document consisting mostly of one deeply nested list could make the view jump on every keystroke. Fixed.
• When using macOS' dark mode with DEVONthink's General > Appearance > Use dark background for documents setting disabled, the text cursor in Markdown documents was invisible with the Default color theme chosen. Fixed.
• Pressing ↩ in WYSIWYG Markdown left blockquotes and callouts rendered at the incorrect height until the document was scrolled or reformatted, inserted a list marker into fenced code blocks inside lists, and appeared to create a multi-line table cell instead of a new row. Fixed.
• When composing with right-to-left text, pressing ↩ in a list or at the end of the text started the next paragraph in left-to-right text direction. Fixed.
• Text copied from WYSIWYG Markdown and pasted into other applications excluded hyphens and pipes, breaking tables and horizontal rules. Fixed.
• The column widths in WYSIWYG Markdown tables were poorly distributed. Fixed.
• Formatting, e.g., bold or italic, applied to a selection spanning several cells of a Markdown table gathered the text of all selected cells into the first cell, discarding the other cells, instead of formatting each cell separately. Fixed.
• Formatting text in a completely selected cell of a Markdown WYSIWYG table merged it with the following cell. Fixed.
• Text typed into a WYSIWYG Markdown table's header cell lost the bold header formatting after the first character, and italic text in the header row was not in bold type. Fixed.
• Pasting text into a Markdown document via Edit > Paste and Match Style kept the default font when pasted into a section heading or other styled line. This also affected other plain-text insertions such as drag-and-drop. Fixed.
• The Edit > Undo command in the WYSIWYG editor could sometimes incorrectly restore old formatting or do nothing at all. Fixed.
• Certain punctuation in a Markdown document, e.g., in C++, could be misinterpreted as formatting characters and inhibit other formatting, and three or more consecutive delimiters, e.g., +++text+++, produced incorrect HTML code. Fixed.
• A backslash after formatting characters or before a line break in Markdown could break the formatting or ignore the line break in Markdown captures and conversions. For example, a clutter-free capture of the text **HOMETOWN\** would be treated as incomplete formatting. Fixed.
• Saving edits to Markdown documents rewrote CriticMarkup insertion, deletion, and highlight blocks as ordinary underline, strikethrough, and highlights. Fixed.
• Escaped square brackets in Markdown documents could render as a LaTeX formula in previews and syntax colorization. Fixed.
• Images in Markdown documents using absolute links, e.g., /Assets/frog.jpg, were not displayed in WYSIWYG mode. Fixed.
• A WYSIWYG Markdown document containing images couldn't be scrolled to the end until the window was resized. Fixed.
• MultiMarkdown 6 abbreviation definitions, such as [>MMD]: MultiMarkdown, were interpreted as link definitions, turning abbreviation references into broken links to the first word of the expansion. Also, definitions indented by one to three spaces were not recognized at all, and saving the document duplicated their marker. Fixed.
• Markdown code blocks specifying their language in an attribute list, such as {.python #cell} or {r, echo=FALSE}, lacked syntax highlighting in previews and exports. Raw blocks such as {=html} were also incorrectly assigned a language. Fixed.
• Images stored in the database or inside a .textbundle file were missing when exporting Markdown as LaTeX and PDF (LaTeX) documents. Fixed.
• Editing inside a fenced code block of a Markdown document with syntax highlighting redrew the whole block on every keystroke, which could make the insertion point jump while deleting text in long blocks (e.g. JSON). Fixed.
• Converting a web page to Markdown dropped text that stood directly inside a block quote next to a label or other formatting, e.g. the abstract on an arXiv page came out as just > Abstract:. Fixed.
• Annotations without a color of their own displayed as yellow when splitting, merging, or saving PDFs on macOS 27 Golden Gate. Fixed.
• PDF notes and other annotations added or changed after opening a PDF document were damaged when saving on macOS 27 Golden Gate. Fixed.
• Text and buttons of a PDF Note annotation ignored the note's color, making them unreadable on macOS 27 Golden Gate. Fixed.
• Text-based annotations rendered pixelated and caused a slowdown and file growth when saving on macOS 27 Golden Gate. Fixed.
• PDF text annotations could cause a random crash while being drawn on macOS 27 and could lose their text in the saved document. Fixed.
• macOS' annotation format picker at the top of a PDF annotation's context menu didn't work on macOS 27 Golden Gate. Fixed.
• Changes made to the table of contents of PDF documents weren't saved on macOS 27 Golden Gate. Fixed.
• If a PDF document was closed or replaced, e.g., via sync or an automation, while the mouse button was pressed, DEVONthink could crash. Fixed.
• Thumbnails of PDF documents added by other applications, e.g., DEVONagent or the MCP server, were not always opaque. Fixed.
• When converting web pages to Markdown, the links of citation markers could be lost, words merged, tables broken, and stray characters left behind. Fixed.
• Dates in RSS articles were sometimes parsed incorrectly. Fixed.
• Clipping or downloading web pages could stall until DEVONthink was relaunched, leaving only a bookmark. Fixed.
• Clutter-free clipping of Substack and similar sites could omit an article's title block or certain leading images. Fixed.
• Images in clutter-free clipped pages could appear distorted. Fixed.
• Web pages with the article wrapped in a paywalled container could only be clutter-free clipped as the headline and teaser links. Fixed.
• Web pages having containers with names coincidentally matching advertising-related terms, e.g., "thread-items", could be ignored when clipping. Fixed.
• Clutter-free clipping and the Reader view dropped section headings that were wrapped in anchor links. Fixed.
• Labels used for screen readers incorrectly appeared above the headline in clutter-free captures. Fixed.
• The Reader view displayed the author and date above the headline instead of below it and could show cookie consent panels instead of the article. Fixed.
• The user agent of web views didn't work with Cloudflare, and images from sites protected by Cloudflare did not load in the Reader view. Fixed.
• No divider separated the item list and view/edit pane in widescreen mode on macOS 15 and later. Fixed.
• In the icon view with details, labeled items in the trash weren't struck through, and labeled bookmarks weren't underlined. Fixed.
• Detached popovers could not always be moved or resized correctly on macOS 27 Golden Gate. Fixed.
• Clicking and holding an item in the Navigation or Editing bars did not open its menu on macOS 26.7 Tahoe and later. Fixed.
• Clicking the empty area next to the Navigation or Editing bars could trigger a hidden tool. Fixed.
• In case of a visible toolbar search without results, several user interface items, e.g., File > Update (Indexed) Items, could be disabled instead of using the sidebar selection to determine their status. Fixed.
• Opening a Reading List item in a new window always opened it in a background window. Fixed.
• When DEVONthink launched or a workspace was restored, the Info > Generic inspector opened instead of restoring the last used inspector. Fixed.
• Markdown annotations in the Info > Annotations & Reminders inspector and the Take Note view of the Sorter didn't always work well in dark mode. Fixed.
• The Document > Links and Mentions inspectors exhibited visual glitches. Fixed.
• Validation of the File > Download Pending Files command was slow when many Markdown documents were selected. Fixed.
• For Markdown documents using right-to-left text, the Format > Alignment > Writing Direction menu incorrectly reported Standard. Fixed.
• The Navigation and Voice Note tabs of the Sorter were incorrectly sized when opened on an external monitor. Fixed.
• The Sorter could close after creating a group using the Location popup menu on macOS 27 Golden Gate. Fixed.
• Shortcut recorder fields, e.g., in the Sorter settings, could display incorrectly after switching between light and dark appearances. Fixed.
• Example item links, e.g., x-devonthink-item://uuid, in documents were indexed as outgoing item links. Fixed.
• After downloading DEVONthink via Safari, the application could prompt to be moved to the Applications folder even if it was already installed there. Fixed.
• Opening HFS+ encrypted or revision-safe databases on macOS 27 Golden Gate asked for the password a second time after canceling. Fixed.

Fixed (Pro)

• In the AI > Library settings, editing a skill's attached JavaScript for Automation (JXA) script opened it in the default viewer instead of the default editor. Fixed.
• Importing two or more .dtSkill files in the Finder by double-clicking them could crash DEVONthink if the AI settings were open. Fixed.
• OCR could fail for PNG screenshots created on macOS 27 Golden Gate. Fixed.
• Applying OCR or transcribing PDF documents containing pages with unreadable content could cause exceptions. Fixed.
• Tool calls for creating and moving items could incorrectly report a database as read-only when neither MCP tools nor skills were used. Fixed.
• The extract_record_visuals MCP tool reported an error instead of an empty result when no existing page was matched. Fixed.
• The extract_record_content MCP tool could stop processing long PDF documents, omitting pages and preventing content on them from being found. Its pages argument could return the complete document, and its query argument was ignored for plain text, Markdown, and bookmarks, with non-matching filters returning the entire document instead of an empty result. Fixed.
• The is_running MCP tool claimed that DEVONthink was not running whenever the server itself was started outside the current session. Fixed.
• Failed HTTP requests of DEVONthink's MCP server incorrectly returned an "OK" status instead of the error. Fixed.
• The AppleScript command get chat response for message… as "JSON" could fail with OpenAI's Responses API if the prompt didn't mention JSON. Fixed.
• Estimated chat costs for OpenAI's GPT 5.6 and GPT 6 models omitted the surcharge for writing entries to the prompt cache. Fixed.
• Models served by OpenAI-compatible servers that strictly validate tool definitions, e.g., Cerebras or xAI, had tool calling disabled. Fixed.
• Requests denied by Claude's safety classifiers returned an empty response instead of the error. Fixed.
• The vision quality from the AI > Chat > Usage popup didn't work with Gemini 3 models. Fixed.
• Gemini's estimated costs were higher than the actual cost. Fixed.

Fixed (Server)

• The web server could crash if it received a request at the same time the connection closed. Fixed.
• Documents with certain characters, such as typographer's quotes, emoji, or Asian characters, downloaded from the web interface with a broken filename. Fixed.
• Logging into the web interface via HTTP authentication failed if the web server's name contained typographer's quotes. Fixed.
• Stalled server connections didn't always close within two minutes of inactivity. Fixed.

Removed (Pro)

• We removed dedicated support for outdated open-source models, e.g., Llama 3 and 4, Phi 4, and Mistral Nemo.
• We removed Meta's Muse Spark contributor model, GPT 5.3 Codex, and DeepSeek V4 Pro from OpenRouter.
• We replaced several OpenRouter models: Muse Spark 1.2 with Muse Spark 1.3, Qwen 3.7 Flash and Plus with Qwen 3.8, DeepSeek V4 Flash 0731 with DeepSeek V4.1 Flash, Xiaomi MiMo 2.5 and MiMo 2.5 Pro with MiMo 2.6, SpaceXAI Grok 4.6 with Grok 4.7, and Z.ai GLM 5 Turbo and GLM 5V Turbo with GLM 5.3.