DEVONthink for Mac

DEVONthink 4.4 turns the AI assistant into a true co-worker. A new AI library lets you create, store, and sync roles, prompts, and skills that can trigger your smart rules, batch processes, and scripts on demand. In addition, the Chat assistant can use DEVONthink's MCP tools directly and improved prompt caching makes multi-step tasks noticeably faster and much cheaper. Of course, we have also updated to the latest models of all major AI providers.

Markdown takes the second spotlight with WYSIWYG support for live tables and MathJax equations, a new syntax colorizer, custom color themes, and direct export to LaTeX or typesetting as PDF via pdflatex. Web clipping is also faster and more accurate, using a new, local Markdownifier and Readability engine. The new Reader view lets you read web content distraction-free.

Finally, encrypted and revision-safe databases use modern APFS images on macOS 26.0 Tahoe and later, and the help viewer's chat assistant is available for free in all editions.

NOTE: We named this release after one the Moon’s most prominent craters, Tycho, itself named after the Danish astronomer Tycho Brahe (1546–1601). Working without a telescope, Brahe compiled the most precise astronomical observations of his era — a dataset so exact that his assistant Johannes Kepler later used it to derive the laws of planetary motion. His work is a reminder that rigorous record-keeping is itself a form of discovery, one whose value can outlast the observer who created it.

NEW

• DEVONthink adds two new export commands to the File > Export menu for exporting Markdown documents to native LaTeX or PDF. LaTeX export resolves WikiLinks and transclusions, carries over title, author, and date metadata, and copies referenced images alongside the exported document; PDF export requires a local installation of MacTeX or the smaller BasicTeX, and merges multiple selected Markdown documents into a single PDF.
• DEVONthink adds a new Reader view that displays a clean, sanitized read of web content like bookmarks, HTML documents, and web archives that contain a detectable article. Open it via View > Document Display > Reader, the Reader button in the Navigation Bar, or ⌃⌘A, then capture the page via the Tools > Capture menu, a context-menu command, or by printing.
• DEVONthink uses a new, local engine for converting web content to Markdown. Beyond speed and native compatibility, it also improves navigation, advertising, and boilerplate removal; metadata, embedded media, and paywall detection; forum thread and authorship detection; callouts and footnotes; and MathML-to-LaTeX conversion.
• DEVONthink adds a Color Theme popup to the Files > Markdown settings, letting you choose a theme for both the source and preview modes of Markdown documents. The Automatic option uses the Markdown font and plain text colors chosen in the Editing > Format settings.
• DEVONthink adds a Date Attributes Modified criterion to the predicate editor for smart groups, smart rules, and advanced toolbar searches. The search prefix is attributesModified or attributesModification and honors the same date-based prefix operators as other date criteria; in the higher editions, the MCP server and Chat assistant support this property too.
• A new read-only AppleScript property, has text layer, lets you quickly identify PDF documents that need OCR.
• DEVONthink brings a new mark record AppleScript command that adds annotations (highlights, underlines, and strikethroughs) to found text within PDFs with text layer, rich text, Markdown, and HTML-based documents like formatted notes. The MCP server supports this too.
• DEVONthink adds new application scripts for use with MailMaven: Add attachments to DEVONthink, Add message(s) to DEVONthink, and Add message(s) & attachments to DEVONthink. Importing attachments requires full disk access.

NEW (Pro)

• DEVONthink adds an extensive new AI > Library settings pane: a repository of roles, prompts, and skills you can use in the Chat assistant and in smart actions. It ships with ten built-in, modifiable functions, each with its own settings and a keyword (slash command) for quick access, and you can create your own manually or with AI assistance. Skills can call scripts (including AI-scripted ones) and trigger existing automations like smart rules, batch processing, and Script with Input/Output actions. Your library syncs via iCloud across your Macs and can also be exported as a package to share. Skills require AI models capable of using MCP tools.
• DEVONthink adds an Allow MCP tools option to the AI > Chat settings and the Chat assistant, giving the AI access to DEVONthink's MCP tools while you chat. The tools honor your other assistant settings and require a model that supports tool calls with a context window of at least 64,000 tokens (currently ChatGPT, Claude, Mistral, Gemini, and Qwen 3.x and Muse Spark via OpenRouter). DEVONthink caches requests to minimize token spending, though MCP calls still take more time and tokens than a plain chat.
• DEVONthink adds a Copy JSON button to the AI > MCP > Install Connectors settings to copy the stdio configuration for use in third-party applications, and a similar Copy Server JSON Config command in the MCP Menu Extra for the HTTP access configuration.
• Tooling-capable AI models can retrieve a PDF's page count in the Chat assistant, and the AI tool that extracts record contents accepts a new pages parameter to target specific pages or a page range. Available to the MCP server, the Chat assistant, AI smart actions, and the get chat response for message AppleScript command, this lets you process large PDF documents in smaller chunks and stay under the context window while reducing cost and processing time.
• A new AppleScript command, perform chat skill, runs an existing skill and returns its results — handy for running skills on a schedule via smart rules. Skills that require confirmation can't be run via scripting.
• A new tool retrieves a PDF document's table of contents in the Chat assistant, the MCP server, smart actions, and the get chat response for message AppleScript command.
• DEVONthink supports Meta's Muse Spark 1.2 via OpenRouter.
• DEVONthink supports macOS 27's fm serve server as an OpenAI-compatible provider.
• DEVONthink supports Google's Nano Banana 2 Lite for fast image generation, though HD quality isn't available.
• DEVONthink adds prompt caching for Mistral AI and Claude models served via OpenRouter.

Improved

• We significantly improved the appearance and reliability of the DEVONthink Markdown parser and renderer. Beyond the new color theme option, the WYSIWYG source now more closely matches the rendered preview, and the WYSIWYG view supports live-edited tables and MathJax equations, not just images and links. DEVONthink also natively renders GitHub-style emojis.
• You can now copy multiple selected ranges in the Markdown WYSIWYG editor.
• Several inspectors use the new parser for consistency, including Info pane > Properties, Content > Table of Contents, Document > Annotations, Attachments, and Links.
• DEVONthink produces more naturally readable mathematical output when converting Markdown documents with equations to plain text.
• Markdown and HTML documents preview more responsively in Side-by-Side view.
• DEVONthink wraps long lines in code blocks within the Markdown preview and Reader view, instead of letting them run past the visible width of the pane.
• The default stylesheet set in the Files > Markdown > Style Sheet setting no longer overrides DEVONthink's internal stylesheet, retaining more of the built-in styling when you edit just a few CSS elements.
• DEVONthink clips web content directly from the loaded WYSIWYG page, more reliably stripping unwanted elements like cookie banners while retaining important elements such as images and scripts, including on sites like YouTube or X.
• When clipping, DEVONthink uses browser-supplied HTML (e.g. by bookmarklet, extension or via AppleScript) more reliably by keeping pages from refreshing or forwarding themselves.
• If you've viewed and logged into a paywalled or authenticated site in DEVONthink, the app can reuse the stored cookies when you clip from that site externally, e.g., via our browser extension, increasing the chance of a successful capture.
• DEVONthink better converts inline code in web content to Markdown.
• DEVONthink lets you toggle the Clutter-free layout when clipping web content as Markdown via the Sorter and the create Markdown from AppleScript command.
• When composing a Markdown note in the Sorter, DEVONthink imports images you drag and drop into it, if the Files > Markdown > Import images to group option is enabled.
• Text annotations added to rotated pages of PDF documents are horizontal now, following the current rotation of the page instead of its original orientation. Their default size, their automatic resizing to fit the text, and their resize handles follow the rotation too.
• In a sheet created with the Tools > Metadata Overview command, DEVONthink defines the Item Link column as a crosslink instead of a URL.
• You can select and copy content in cells within read-only sheets, e.g., locked or revision-proof databases.
• DEVONthink creates encrypted and revision-proof databases as APFS disk images on macOS Tahoe and later instead of using the older HFS+ format. Note that you cannot open databases using the new APFS encryption on older versions of macOS.
• DEVONthink unmounts encrypted and revision-proof databases more reliably.
• DEVONthink supports the matches (:) operator, with quoted or unquoted text, when searching custom metadata, e.g., mdstatus:"To Do".
• Going to a page, line or row of a sheet no longer depends on the system's number formatting, both 1018, 1 018, and 1,018 work now.
• Multiple selected items with thumbnails display as a fanned stack in the view/edit pane.
• The Info filter supports filtering by Date Attributes Modified.
• We made many minor cosmetic tweaks and fixes to the interface, touching inspectors, the splash screen, the License panel, etc.
• The Databases Finder folder has a new icon on macOS Tahoe and later.
• We updated the layout of the Help viewer and the help pages, including a reordered sidebar.
• DEVONthink makes the Help assistant available for free in all editions and during the trial period. We are providing this functionality as a courtesy, but note that you can no longer change its AI provider.
• DEVONthink retrieves ISBN numbers more reliably via AppleScript's international standard book number property and smart action placeholders.
• We renamed the AppleScript property attributes change date to attributes modification date, keeping the old name as a synonym for compatibility.
• The replicate record to AppleScript command supports replicating to an RSS feed.
• The Export > Listing script runs at least six times faster.
• We improved language detection for documents, making it faster and more accurate.
• DEVONthink processes and indexes HTML content more effectively, removing unnecessary information to improve searches and conversions.
• DEVONthink clears smart group and smart rule caches immediately when you close a database.

Improved (Pro)

• Imprinted documents achieve better compression when saved.
• The Chat assistant and MCP server better handle Markdown documents with empty sections.
• You can file documents and groups created via the Chat assistant by providing a name or path.
• In the AI > Chat settings, DEVONthink updates per-model usage costs after each completed request using current rates, for more precise totals. Costs also reflect the discounted rates for cached data where applicable, and show an estimated percentage of cached input.
• All the AI functions, e.g., the Chat assistant, smart actions, scripting commands, AI popovers, etc., redact your name and phone number, and the phone numbers of your other contacts. The MCP server optionally does this as well.
• Better caching and response filtering reduce costs and increase accuracy in the Chat assistant, AI smart actions, and the get chat response for message AppleScript command.
• We improved error logging and handling for the Chat and Help assistants.
• Local MCP server connections no longer require a TLS certificate, though one is still recommended.
• In the AI > MCP settings, Install Connectors > Codex Desktop & CLI supports the latest ChatGPT application.
• The MCP server can get color label names, more flexibly access the Global Inbox, and better handle UUIDs.
• DEVONthink's MCP server supports the latest version of MCP as well as previous versions. This improves compliance and reliability in connection handling and also prepares the server component for future MCP-compatible applications.
• All MCP tools declare the scope of their actions via annotations. These denote whether an action can access the internet, modify or overwrite data, or is read-only. AI clients that support these annotations, e.g., Claude Code, use them to determine if user confirmation is required.
• We updated Claude Opus and Sonnet to version 5, with a knowledge cutoff of 2026-07-01; Sonnet has a context window of one million tokens and 128,000 output tokens. Pricing hasn't changed, though the new tokenizer uses up to 30% more tokens per request. Claude's prompt caching is much more efficient, making multi-step tasks noticeably faster and reducing token costs.
• DEVONthink supports the latest OpenAI models, GPT 5.6 Sol, Terra, and Luna, each with a one-million-token context window, 128,000 output tokens, and knowledge current to February 16, 2026. The Script assistant uses 5.6 Sol.
• We updated Gemini Flash to 3.7 and Flash Lite to 3.1, both accessible via API and OpenRouter, with knowledge current to 2026-07-01.
• DEVONthink uses the OpenAI responses API and Google's interactions API, supporting the latest models and features, including image tool call results.
• DEVONthink offers improved support for running GLM 5.x and muse-glimmer models locally.
• DEVONthink supports context windows of up to one million tokens for Ollama, LM Studio, and OpenAI-compatible providers, up from 128,000.
• Local AI, Ollama, OpenAI-compatible, and Mistral models support longer tool call cycles.
• DEVONthink offers OpenAI's GPT 5.5 (Pro) via the API.
• The Script assistant supports GPT 5.3 Codex. We updated the OpenRouter models used by the assistant: Claude Opus and Sonnet 5, Gemini 3.7 Flash, Grok 4.5, and Kimi 2.7 Code.
• MCP tools and AI skills support Qwen 3.x and Muse Spark models via OpenRouter.
• Transcriptions without timestamps done via OpenAI use GPT-Transcribe instead of GPT-4o.
• The Script assistant and AI skills produce better AI-generated scripts.

Fixed

• Selecting text in the Chat assistant could sometimes crash on macOS 27 Golden Gate. Fixed.
• Viewing or editing Markdown documents containing lists could sometimes crash. Fixed.
• Markdown documents converted to HTML incorrectly contained a second title header. Fixed.
• MathJax in Markdown didn't always render properly. Fixed.
• A Markdown document with unclosed YAML front matter at the top of the document displayed and converted as if it had no content. Front matter now ends at the first blank line, similar to MultiMarkdown. Fixed.
• In Markdown, bold or italic text followed by a superscript character wasn't formatted correctly. Fixed.
• Line items spanning multiple lines in Markdown documents' to-do lists incorrectly showed the text on the line under the checkbox. Fixed.
• With Syntax highlighting enabled in the Files > Markdown settings, callout titles in the Markdown source didn't display in bold and tinted colors like in the rendered view. Fixed.
• Saving edits to Markdown and HTML documents in side-by-side view forced a full reload of the preview, causing flickering and rescrolling. Fixed.
• When text was selected in the Markdown source while using the Side-by-Side view, the Edit > Set Name As command was disabled. Fixed.
• The Smart quotes and dashes option in the Files > Markdown settings got disabled after restarting DEVONthink. Fixed.
• Saving edits to formatted notes could reprocess styles and images, leading to image scaling issues. Fixed.
• Clipping HTML pages with malformed markup, legacy multimedia content, or self-closing tags could cause the content to be lost. Fixed.
• Clipping to formatted notes could cause random issues. Fixed.
• PDF annotations could get lost or damaged on macOS 27 Golden Gate when editing them or when pages were copied, e.g., by splitting, merging, or converting documents. Fixed.
• Searches beginning with NOT incorrectly counted the first term twice, inflating the item's ranking. Fixed.
• Searches using the NOT AFTER operator didn't match documents whose search term was the last word of the text. Fixed.
• Exact phrase searches, e.g., "a b c", could return similar but non-matching documents, e.g., one containing "a b b c". Also, if a document contained repeated words, e.g., the cat the cat the dog, a quoted search for "the cat the dog" didn't match that document. Fixed.
• Sorting very large groups or smart groups by URL could sometimes cause a crash in recent versions of macOS. Fixed.
• Windows opened via the context menu in the Navigate sidebar or item list would open in the background on macOS 26 Tahoe and later. Fixed.
• The toolbar search field in the Help viewer wasn't vertically centered on macOS 27 Golden Gate. Fixed.
• The Navigation Bar was sized incorrectly when previewing Markdown documents, HTML pages, or web archives on macOS 27 Golden Gate. Fixed.
• When inserting a back link or quote in an annotation file, Asian characters in Markdown section names were converted to Latin. Fixed.
• Documents with text containing the name or alias of another document weren't shown in the Document > Mentions inspector if the first word of the name was repeated, e.g., skill skill prompt didn't match a document named skill prompt. Fixed.
• When the maximum number of versions per document was reached, the next save would incorrectly remove the most recent version instead of the oldest. Fixed.
• You couldn't drag and drop multiple emails from Apple Mail into DEVONthink on macOS 27 Golden Gate. Fixed.
• When trashing a replicant matched by a local smart group, the menu and context menu command was "Move to Trash" instead of Move All Replicants to Trash. Fixed.
• The Scan Name and Scan Text smart actions could cause an exception when processing a document with no text, e.g., a bookmark. Fixed.
• Scheduled smart rules that hadn't run due to having no matched items could cause high CPU usage. Fixed.
• An alias or symbolic link in the Scripts or Templates menus that pointed to an enclosing folder could make DEVONthink freeze and finally crash. Fixed.

Fixed (Pro)

• When browsing certain sites, too many items were added to the Download Manager. Fixed.
• If the MCP server failed to launch on startup, the alert always showed the default 8420 port instead of the one configured in the MCP settings. Fixed.
• Summarizing or tagging large documents with Apple Intelligence could produce frequent "Exceeded model context window size" errors. Fixed.
• Gemini didn't work reliably via OpenRouter. Fixed.

Fixed (Server)

• In the web interface, dates changed in the Data inspector were not saved. Fixed.
• Some Markdown documents in search results couldn't be edited. Fixed.
• The Generic and Data inspectors were disabled when a search result in the web interface was selected. Fixed.

Removed

• We removed MultiMarkdown support and its renderer popup from the Files > Markdown settings.
• We removed the Clutter-free layout option from the Tools > Capture menu, as it's no longer needed with in-application clipping.
• We removed the obsolete iData file plugin (e.g., .idata, .id3).
• We removed a workaround of the List and Icons views in widescreen preview mode on the latest versions of macOS 26.x so that the scrollbar is not always visible anymore but honors the system's setting.

Removed (Pro)

• We removed the outdated O3 and O4 Mini models from OpenAI access and OpenRouter.
• We removed Magistral models from Mistral and OpenRouter access.
• We removed Meta Llama 4 Maverick and Scout, and DeepSeek R1, from OpenRouter.
• We replaced several OpenRouter models: GLM 5.1 with GLM 5.3, DeepSeek V4 Flash 0423 with DeepSeek V4 Flash 0731, DeepSeek V4 Pro with DeepSeek V4 Pro 0813, Qwen 3.6:27B with Qwen 3.8:27B, Qwen 3.6 Flash with Qwen 3.7 Flash, Qwen 3.7 Max with Qwen 3.8 Max, and Grok 4.3 with Grok 4.6.
• We removed the Role popup from the AI Chat settings; specify roles in context instead, e.g., via a skill.