DEVONthink for Mac

DEVONthink 4.3 Herschel introduces our MCP server: A safe, private bridge that lets Claude, Cowork, Codex, and other AI clients work directly with your databases. No scripting or third-party tools required. On macOS Tahoe, we now support Apple Intelligence for light tasks and Apple's faster SpeechAnalyzer for transcription. Other AI improvements include a leaner search assistant, a dedicated tagging model, and web search via ollama.com.

A new Markdown parser and renderer replaces MultiMarkdown 6. It adds callouts, citations, and better compatibility with AI output. Desktop widgets bring the reading list, favorites, and workspaces to macOS Sequoia and later. For DEVONthink Server users, the web interface gains classification, context menu commands for moving, duplicating, and replicating items, and a dialog for opening item links. It also features a much better PDF viewer. As always, reliability and performance continue to improve.

NOTE: We name this release after the periodic comet 35P/Herschel–Rigollet, discovered in 1788 by Caroline Lucretia Herschel (1750–1848). Caroline was the sister of the famous astronomer William Herschel and made significant contributions to astronomy in her own right. She discovered eight comets and numerous nebulae and star clusters, and compiled an influential catalogue of deep-sky objects. Caroline Herschel was the first known professional female astronomer, became the first woman to receive a salary as a professional astronomer, and received the Gold Medal of the Royal Astronomical Society.

NEW

• Replacing MultiMarkdown, our new Markdown parser and renderer supports features like callouts, citations, and tables with captions. It also brings improved CriticMarkup support and is more compatible with AI-generated output. Choose the renderer in the Preview popup in the Files > Markdown settings.
• Supported in macOS Sequoia and later, new desktop widgets let you access your Reading List or Favorites, load a Workspace, or view a tip from our blog.
• View .highland Markdown files used by screenwriters.
• The new URL command, x-devonthink://open-workspace?name=, lets you open an existing workspace.

NEW (Pro)

• DEVONthink gets its own MCP server that provides a bridge between your databases and supporting AI applications, e.g., Claude. Comprising a suite of almost 60 commands, it is built with data privacy and security in mind, controlling local or remote access, redacting private information from being transmitted, honoring all exclusions from AI set on items (or databases), and limiting access to the internals of your database in the filesystem. In addition, when the server is running, a menu bar icon appears with commands for controlling the server, accessing the settings, or copying the URL you need. Requires macOS Sequoia or later.
• Use the new Settings > AI > MCP settings to set up and install the DEVONthink MCP server. The options include settings for privacy and making safe connections to other devices, on and off your network.
• The option Exclude from Chat & MCP in the Database Properties panel lets you easily exclude entire databases from AI access. This is enabled by default for encrypted and revision-proof databases.
• DEVONthink supports Apple Intelligence's Foundation model in the AI > Chat settings and AppleScript commands on macOS Tahoe and later. At this time, its use is limited to simple tasks like renaming, tagging, etc. as the context window is very small and vision is not supported.
• Choose the AI model for automatic tagging functions with the new Chat Suggestions popup in the Files > Tags settings. DEVONthink uses this, e.g., when importing files, using the Data > Tags commands, or smart rules and batch processing actions.
• Added support for LM Studio's reasoning effort (0.4.8 or later).
• You can now use an ollama.com account for web searches in the AI > Search settings.

NEW (Server)

• Added an Open Item Link button to the Action toolbar button and the context menu of the item list in the web interface. This opens a dialog to enter an item link to be revealed.
• Added a Reveal command to the context menu of search results in the web interface's item list, revealing the item and its location in the Navigate sidebar.
• Added Move, Duplicate, and Replicate commands to the context menu in the web interface's item list, allowing you to navigate to and choose the destination for these functions.
• Added a section for classifying items to the See Also inspector of the web interface, including a context menu with related commands.

Improved

• Converting Markdown documents to plain text, e.g., via the Data > Convert > To Plain Text command, strips all Markdown formatting and converts MathJax and DisplayMath into a more readable form.
• Sheets transcluded in Markdown use an underlying Markdown representation for a more consistent appearance across documents.
• The "What's New" window also appears while DEVONthink is in fullscreen mode.
• Copying item links includes a Markdown-formatted link for use with third-party editors.
• You can use an alternative syntax, :==, synonymously for the is search operator, e.g., in created:==2026-04-14.
• Commands like Open Externally use fewer resources, improving DEVONthink's responsiveness.
• Both the digital object identifier AppleScript property and the Digital Object Identifier smart rule and batch processing placeholder use also the URL to retrieve the DOI. The chat assistant and MCP server of the Pro edition benefit from this too.

Improved (Pro)

• AI responses use our new Markdown parser, e.g., in the Chat assistant or the summary and transform popovers. Additionally, the Chat assistant knows how and when it's appropriate to use Markdown callouts (admonitions).
• The Chat assistant, AI smart actions, summarize and transform functions, AI-related AppleScript commands, and the new DEVONthink MCP server, redact or anonymize email addresses, links, credit card numbers, phone numbers, authentication tokens, and labeled secrets in the information sent to LLMs.
• The Chat assistant displays images returned by Perplexity Sonar models.
• The Help assistant uses fewer tokens and is faster and more reliable.
• "Exclude from Chat" in the Generic inspector is now Exclude from Chat & MCP.
• Local Apple SpeechAnalyzer replaces the "Local Apple Speech" option in the AI > Transcription settings on macOS Tahoe and later. This utilizes Apple's new SpeechAnalysis framework which processes audio locally and up to five times faster. Additionally, it will automatically download and install resources available for the specific language being transcribed.
• Researching PubMed and arXiv via the Chat assistant is more reliable.
• If AI requests the text contents of a document, DEVONthink provides a Markdown representation for better access to things like links and styling. This applies to RTF(D), HTML pages, formatted notes, and web archives. (PDF and EPUB documents are already supported). It also improves responses from many AI features, e.g., the Chat assistant and the get chat response for message AppleScript command.
• Batch tagging is more reliable with low-tier models, e.g., Magistral.
• The default role for AI-related features, e.g., the AI smart actions, get chat response for message AppleScript command, etc., includes the date, weekday, and current time.
• If the digital object identifier (DOI) or international standard book number (ISBN) is in the text, metadata, or content of a document, it is available to AI functions like the Chat assistant, Chat - Query smart action, or get chat response for message AppleScript command without processing the document further.
• Tagging documents with AI retains the spelling when Files > Tags > Use only existing tags of a database is enabled.
• An improved definition of DEVONthink's search syntax makes database searches via the Chat assistant, Chat - Query smart actions, and the get chat response for message AppleScript command more precise and reliable.
• Claude Opus is now at version 4.7, both via API and OpenRouter. With a knowledge cutoff of January 2026, it supports up to one million input tokens and 128,000 output tokens as well as high-resolution image processing, up to 3.75 megapixels. The Sonnet models also support up to one million tokens now.
• Updated OpenAI's GPT models to version 5.5, supporting a context window of one million tokens and a knowledge cutoff date of December 1, 2025, but at a much higher cost. Also, the internal names of these models no longer reference the version, ensuring future updates don't adversely affect scripts, templates, or smart actions.
• Updated Gemini 3.1 Flash (Preview) to Gemini 3.5 Flash, both via Google and OpenRouter. Note that the new model is three times as expensive as the former one.
• Gemini 3.1 Flash Lite is available via API and OpenRouter.
• Updated the Mistral Small model to version 4 and Mistral Medium to 3.5, both with a context window of 256,000 tokens and reasoning capabilities. However, these models are much more expensive than the previous ones.
• Qwen 3.5, Gemma 4, and versions 3 and 4 of Mistral Small are better supported locally and via OpenRouter.
• Improved token usage when using Perplexity for Web & Wikipedia searches.
• Updated the Flux image generation models to Flux 2 Klein, Pro, and Max. They support standard and HD quality and image editing, but not styles.
• Nvidia's Nemotron 3 is better supported.
• Xiaomi models are available in OpenRouter.
• Updated many models in OpenRouter: Grok 4, replaced by Grok 4.2. Mistral Small 3.2, replaced by Mistral Small 4. Gemma 3, replaced by Gemma 4. GLM 4.6 and 5, replaced by GLM 5.2. GLM 4.7 Flash, replaced by GLM 5(V) Turbo. Xiaomi MiMo-V2-Flash/Omni/Pro, replaced with MiMo-V2.5 and MiMo-V2.5 Pro. DeepSeek 3.2 and DeepSeek 3.2 (Speciale), replaced with DeepSeek V4 or DeepSeek V4 Flash. Kimi K2 2.5, replaced by Kimi K2 2.6. All Qwen models, replaced by various 3.6 models.
• OpenAI's GPT-OSS models are better supported via LM Studio or Ollama local.

Improved (Server)

• A better PDF viewer in the web interface supports a table of contents, thumbnails, and searching in all browsers. Additionally, toolbar search terms are automatically used when selecting a PDF document in the search results.
• The context menu in the web interface's item list is better supported on mobile devices.
• Double-clicking a group in the search results of the web interface's item list now opens the group.
• DEVONthink's self-signed certificates generated in the Server > General settings use even stronger 2048-bit encryption.

Improved (Sync)

• Synchronizing RSS feeds less frequently duplicates articles.

Fixed

• Converting Markdown documents to paginated PDF documents could sometimes fail. Fixed.
• Opening braces with no matching closing braces could break Markdown transclusions. Fixed.
• Converting images to PDF or merging PDF documents on macOS Tahoe could produce a document with only white pages. Fixed.
• Scrolling or selecting pages in a PDF document could mark the document as modified. Fixed.
• On rare occasions, annotating PDF documents could cause an exception. Fixed.
• Multi-page TIFF files couldn't be viewed or converted on macOS Tahoe. Regression. Fixed.
• Web windows opened without a record, e.g., via AppleScript, always opened at the same size with the Inspectors open. Fixed.
• Scrollbars in Widescreen mode didn't behave properly on macOS Tahoe. Fixed.
• Resizing document windows after switching to side-by-side mode didn't work correctly on macOS Sequoia and later. Fixed.
• Group tags didn't appear correctly in inactive tag fields. Fixed.
• In tagging fields, the popup arrow on hovered tags didn't appear in the correct position. Fixed.
• Depending on the date, time, and current locale, the contents of the Log window's Date/Time column could appear clipped. Fixed.
• Tooltip and click positioning in the Editing and Navigation bars were affected by a window's width on macOS Tahoe. Fixed.
• When using a Markdown document in the Annotation field of the Annotations & Reminders inspector, the Format > Add Link and Format > Edit Link commands weren't supported. Fixed.
• The Content > Table of Contents inspector didn't display correctly with Markdown documents using more than three backticks to mark fenced code blocks. Fixed.
• Clearing the color of groups and tags from the Generic Info inspector or Info popover didn't sync between devices. Fixed.
• On startup, a failure to open a database showed an alert that could freeze the application on rare occasions. Fixed.
• The JavaScript command addReminder didn't use the correct scheduling. Fixed.

Fixed (Pro)

• Transcribing PDF documents on demand, e.g., via Data > Recognition or automatically, produced inconsistent results. Fixed.
• Empty black or white PDF pages weren't ignored as input, leading to unnecessary AI time and token usage on macOS Tahoe. Fixed.
• Blockquotes in the Chat assistant weren't styled. Fixed.
• Text pasted into the Title field of the Imprinter settings incorrectly added a % character. Fixed.
• Adding or deleting a Font placeholder in the Title field of the Imprinter settings didn't update the Font selector popup. Fixed.
• Inserting a placeholder before an existing placeholder in the Imprinter settings replaced all the text following the new one. Fixed.
• Custom metadata definitions added via AppleScript could cause sync issues. Fixed.
• If DEVONthink was not running, the first scan from third-party software launched the app but didn't run OCR on it. Fixed.

Fixed (Server)

• In the web interface, PDF documents couldn't be viewed in Safari. Fixed.
• PDF documents viewed in the web interface on an iPad only displayed the first page. Fixed.
• Markdown documents didn't display images using relative links in the web interface. Fixed.
• The table of contents for EPUB files didn't work. Fixed.
• The arrow indicating the item list's sort direction was the opposite in the web interface. Fixed.
• The user icons in the Server > Users settings didn't display correctly when launching in dark mode and switching to light mode while the settings were open. Fixed.

Removed (Pro)

• Removed the Remote Apple Speech option from the AI > Transcription settings on macOS Tahoe and later.
• Removed OpenAI's GPT 4.1 models from API access and OpenRouter. We've also removed GPT 5.1, GPT 5.2, and GPT 5 Nano and Mini from OpenRouter as they have been superseded by newer models.
• Removed Claude Opus 4.6 (Fast) from OpenRouter.