You can do this right now with Claude Cowork without any MCP at all if you use an LLM which is good at Applescript such as Claude 4.6 or Opus 4.7.
For some documents this is doable with Applescript alone. For others it may be necessary to create an OCR index which adds page numbers to the text layer. Claude Cowork can add this for you by installing one of several open source packages which can do that.
I have a Skill set up which creates hyperlinks to source documents and uses the Devonthink URL parameter that allows highlighting specific text within the document.
There are also now several unofficial DT MCP servers - these generally work faster and more consistently than writing random Applescript. If there is a feature such an MCP server does not have, Claude can help you modify the open source MCP server code.
Alternatively Claude can create a second “personal” MCP server whcih adds features not present in an existing MCP server - you can use multiple MCP servers simultaneously.
I have not experimented at all or really researched the capabilities, but I have long wanted to use AI to search my documents easier. I tag almost everything and the basic search (especially on iOS DTTG) never includes or prioritizes tags in to the search (and I am usually trying to search too quickly to want to use the tags:option. I have never understood why basic search, especially when sorting by relevance doesn’t include tags. And it’s been asked for before: Search for tags in DTTG - #5 by halloleo
Of course! I completely understand where you’re coming from, and I appreciate you taking the time to respond.
To bring it back to the spirit of this thread… what we might actually do with the new MCP capability. I was picturing a smart search tool that feels effortless. Right now, when users want to search for something specific, it already takes more effort than it should to surface truly relevant results. And when we try to weave the tags feature into that process, it often pulls in quite a bit of noise instead of sharpening the focus. I’ve searched for the same thing multiple times to show my students relevant videos I have saved and unless I formulate my search query exactly correctly, I keep coming up with the same problem even though I know the tags have the data. Sometimes I resort to renaming the file just because I know that I searched for it that much more often.
Imagine how smooth and powerful it would feel if someone could simply type their search terms and let “sort by relevance” intelligently factor in matching tags—so the results that matter rise to the top naturally. The forum post I referenced is really just highlighting that same everyday frustration many of us run into.
I’m confident that building something like this with MCP could make a noticeable, positive difference for a lot of DEVONthink users. Curious to hear your thoughts on what might be possible here.
Oh, and to be clear, the search in DEVONthink seems to be much better on the relevant based results like I’m discussing, but it’s particularly bad on DTTG which is why I might play around with the MCP interface.
Or custom metadata? Labels? EXIF data? Location?
There’s no reason why tags should be treated any different from other metadata. Using the appropriate operators ensures that you find what you’re looking for. And prevents search results encompassing too much.
I was picturing a smart search tool that feels effortless. Right now, when users want to search for something specific, it already takes more effort than it should to surface truly relevant results. And when we try to weave the tags feature into that process, it often pulls in quite a bit of noise instead of sharpening the focus…
Imagine how smooth and powerful it would feel if someone could simply type their search terms and let “sort by relevance” intelligently factor in matching tags—so the results that matter rise to the top naturally.
You are imagining a prescience that doesn’t exist in AI… well DEVONthink’s own internal AI has been called that or “magic” It’s easy to imagine many things but our flights of fancy can quickly surpass what technology is capable of.
I don’t know what subject you teach but I’d imagine you correct papers: telling students to clarify things, be specific, be concrete, etc. If you type video, engineering, 2024, how is the application supposed to know the context of those terms mean when you haven’t specified it? This is especially bad form since it sounds like you’re searching for parts of names as well as tags, maybe other attributes. You should be specific and clear in your searches, just as it’s good to be specific in clear in other forms of communication. This leads to better, more positive results in many facets of life.
This is astounding. I have a skill in Claude that’s now tagging the 5,000 items in my inbox I’ve neglected. I feel sorry for them. It’s a dream come true for me, and a menial hell for Claude. I hope they don’t crawl this page and link it to me.
Did you ever try the older AI offering by DT to tag your documents automatically? It’s described in the manual under “AI and Tagging”.
I’m wondering if the results of the two approaches are comparable.
Some people here might say you could just ‘write a script’. They fail to realize how useful ai is for the maintenance portion of note taking. You have the custom metadata you can fill till your heart’s content. Things which seem good on paper but one never does in practice can now be automated with ai.
I did, but I couldn’t easily force it to comply with my tag hierarchy and it was incapable of suggesting new tags - and @anonny, I did in fact write a shonky AppleScript (that I never updated past DTP2 ) that matched record keywords against the tag hierarchy - but context can make fools of us all, and it also couldn’t suggest new tags - Claude will read the content and identify thematic tags that are not in the text. Other than draining my usage limits, the Claude Skill has knocked off 200 items this morning, which is about as many as I’ve done in about 7 years, plus I’ve got a bunch of new tags and it’s helpfully pointed out items which have only captured cloudflare interstitials!
Oh! You organize through mindmaps, do you? Xmind MCP Server - Xmind or if that’s not to your liking mermaid maps.. Heck, xml so pretty much whichever mindmapper you like!
Of course, the important part is thinking what you want with precision, but as to the moving the mouse around and clicking, and all the tedious gruntwork, your ‘clanker’ can do that drudgery!
DTTG4. Not quite, I am a helicopter flight simulator instructor and I often pull up case study videos or other relevant files/videos I have saved for students. I also have quite a bit of experience in coding as a computer engineer, and always frustrated that DTTG4 search never finds what I am looking for without having to specify tags or other metadata. BUT, the same exact search in DEVONthink 4.3 Pro finds the files easily when sorted by relevance.
In fact, I just tried a search for two or three keywords and the result I am looking for is #1 by relevance in DEVONthink with 63 results, and I can barely find the same result in DTTG4 (which returns 106 results from the same database.) I understand DTTG4 is a mobile application, but as the primary interface to my data when away from home, I hope the search gets more intelligent after every update.
So, when you asked what I would do with an MCP server, I figured I would write my own mobile application to search my files.
Bear in mind, DEVONthink and DEVONthink To Go are different applications on different platforms with differing underlying technologies though we continue to bring parity between the two where possible. Also, DEVONthink’s internal AI plays a part in its search functions, establishing relevance in a way our mobile app can’t at this time.
I figured I would write my own mobile application to search my files.
You’re going to write your own bespoke MCP-capable mobile application? That’s not something achievable with DEVONthink’s MCP server alone.
I do not know what @CGDaveMac has in mind or whether he has hard-core mobile developer skills.
But even if he does not building an iPhone app to search Devonthink using the MCP would be pretty straightforward via Vibe Coding with Replit. It even helps packaging it for installation on iOS either privately or via the app store. I have been thinking about something similar and I suspect it could be done in half a day on Replit.
It’s easy to imagine many things but our flights of fancy can quickly surpass what technology is capable of.