As we always appreciate insights into the how and why people use our apps… added to the buzz about MCP servers…
If DEVONthink had its own MCP server…
And as always, thanks for your participation! 
I would use the MCP in Claude Desktop/Cowork to create Markdown documents summarizing a conversation, and save artifacts like charts or spreadsheets, things like that. However, I would also use it a lot in Claude Code, and so would prefer a CLI tool that matches the MCP functionality to save token use, as MCPs can blow up your context window even if you aren’t using them for the current session. A CLI with either a detailed --help output or a skill file would probably be plenty for CC to use
An MCP we could use on iOS devices would also be great. But because DT4 is local first, and sync is just syncing the DB files via iCloud/Dropbox/WebDAV, I don’t know this would be possible. There would need to be some managed service that would allow access to our data. Off the top of my head, one possible way to achieve that would be if the desktop version could be accessed via proxy thru an OAuth secure link via a service managed by DT. That would require DT4 to be running and the Mac it’s running on to be always on, and that adds a security risk. So that may not be practical
Thanks for the first response!
I’m curious what you’d do with it in Claude Code.
My MCP experience is limited — the one integration I actively use is Claude with Fantastical for schedule, which works surprisingly well.
For DEVONthink specifically, I see two core use cases:
Write — feed Claude sources (links, webpages, PDFs) and generate documents directly in DEVONthink in whatever format I define.
Though the existing AI Chat inside DEVONthink already can do that…
Retrieve — cross-database search and synthesis. Finding conceptually related documents across separate databases.
these are the use cases I can imagine for myself.
Thanks for this thread!
And this has long been possible with DEVONthink’s internal AI in the See Also & Classify inspector 
But thanks for the thoughts!
A few quick things: I often do extensive planning in CC, and like to save those plans somewhere, particularly if my coding will span many sessions. Typically I save a Markdown file so I can point future CC sessions to it. Or I may have a series of screenshots for setting up some service in the AWS console that I save in DT4 so I can automate that. Right now I have to drag them in from DT4 into the CC session, or copy/paste them in. So with an MCP or CLI, I can tell CC where all of those artifacts live and also have Claude update documents as needed, which is also useful to remember what I had worked on 6 months ago
what would we do with it? let’s think about the strengths of DEVONthink: linking, and metadata.
Obsidian: linking is poor once you step outside of its ecosystem.
notion: similar, and it’s actually hard to extract yourself from their ecosystem.
DEVONthink is like a folder, and it has a lot of things you can do with it.and, as long as you have a mac, you’re able to keep your things as organize as you’ve made them.
The MCP removes the friction.. I need to create another file, and think which files does that one file link to: with an mcp you can say, I have this link: create a link with the top 3 links so that it’s a more ‘stable’ link, and It will ‘walk’ the linked files of the linked file to identify which are actually ‘coupled’ and which are not…
I’d want to use it to retrieve documents, mostly. I sometimes have models write AppleJS to get stuff from DEVONthink. An MCP would be better for that and it could educate the model about DEVONthink’s own search, classifier, related word lists etc.
I’ve been playing with Claude CLI accessing a a bunch of MCPs offering up conventional databases (of very specialized dictionaries, of manuscript catalogues) and also of semantic databases (of multilingual texts). It’s truly fascinating the sort of unique threads I’m able to tug on across these things. A DT MCP would then offer me an interface to query both a large DB of texts I’ve assembled over the years (though I might just move those into a ChromaDB anyway), and more interestingly, my own notes on things. Again, it’s the slicing/sliding across previously silo-ed repositories that has been truly generative.
I would primarily use it in Claude Cowork. Over the past few months, I have been using Cowork extensively for a wide variety of tasks across both personal and business projects. During this time, I have tweaked its instructions to improve its persistent memory file hierarchy. Though I’ve made some decent improvements, I’m running into an issue where storing too many documents is making data retrieval less effective.
For documents such as contracts, agreements, PDFs, etc., a DEVONthink MCP would allow the AI to search and retrieve directly from DTP. Currently, these documents exist in two locations (DEVONthink for storage and Cowork’s file hierarchy for processing) because Cowork can’t access my preferred storage location. They also need to be converted to MD for less token use.
In many of my Cowork sessions, documents get created based on the results of the session, and instead of trying to figure out where to store them in the File hierarchy, they could be saved directly to DEVONthink. Then, it only needs to add the location to its internal file mapping.
But I think the biggest win would be the “better iteration” of the Kerpathy Wiki that people are playing around with. Using DEVONthink as the repository for documents - which is where they should live - and being able to access the data with the MCP would be a stronger system.
Hi, would use MCP (or better CLI) especially with Claude Code. Use cases I would think about:
I’m answering this as someone who doesn’t currently use LLMs (for various reasons) but who has been trying to keep up with developments (feels like it’s necessary even if I’m not engaging directly yet!).
I don’t know that I actually have a use case for this yet, but I’ve been thinking about OpenClaw and how some LLMs allow you to build up a sustained “relationship” with an LLM that is unique to you (the way you like to be “spoken to”, particular ways of working, previous decisions you’ve made), and how interesting that might be if the LLM was plugged in to my main database (which is where all my research and reading lives). It would open up opportunities for dialogue and “discussion” on materials I’m already familiar with, while also hopefully limiting the scope to real citations (a big issue!). It would also then be able to act as a research assistant (“I read something about the North American wolf using crab traps, can you find it and also pull up my notes about the observations? Actually can you compile a list of all mammal tool use papers that I’ve made notes on?”). An MCP server takes us a step closer to this kind of “me and my robot sidekick” scenario.
(But incidentally if DT implemented a way to do this at a database or group level directly we wouldn’t need an MCP server. And wouldn’t it be cool if we could get to a position where the LLM is local and is running off files just on your device, with an accrued knowledge of who you are and how you like to work!)
I would use it to create detailed structured outlines of long documents.
With that in mind - one addition to DT4 Applescripting would be very helpful…
Presently it is possible to initiate a Search via Applescript. However it is not possible via AppleScript to populate the in-document search inspector. That is a really useful inspector because it shows search hits in context.
Any chance that could be added to DT4 Applescript? If yes then it would be helpful in turn if that scripting feature were available via the MCP server.
I’m interested in researching AI and its use in conjunction with engineering, which is my field. I see opportunities for use as long as you understand the strengths and weaknesses of AI (LLM/agents). I believe I see areas where standard engineering practices can help AI be more successful than it is today as well as how AI can facilitate issues we have in engineering, so I’m looking for synergy while keeping the engineer in the loop for the critical thinking -guidance/thinking/reasoning/creativity/verification. Much of engineering requires specific, rigorously controlled, accurate, and highly technical document creation and maintenance. I’m testing how such could be produced as a collaborative AI/engineer effort through the use of current engineering practices and guided AI to address the weaknesses of each with the strengths of the other. So at the outset, I’ll probably use it as I do now for RAG, but I’d like to see if I could next replicate experiments where I’ve iteratively created engineering documents with the AI that required considerable overhead/interaction. I think the “Cowork”-style software would be a good development and test environment and would want Devonthink to be the repository / tool for additional analysis.
I use Devonthink as integral for my research - literature repository/notes, engineering references, ai experiment conversation use and storage, project documentation, etc. I’m interested in the MCP server both to better understand the use of MCP and to integrate with external open source Claude Cowork alternatives that run locally. At this time I’m not interested in using cloud based AI for any sensitive use and willing to pay the performance penalty for such. I have quite a few research questions and experiments I’d like to do and I would like to see how Devonthink would best fit to those needs. The external AI integration with Devonthink has been great and it seems MCP would be the next step in that path,
Wow! At this moment, an MBP M5 Max with 128 GB of RAM is leaving China… I’m in discussions with Gemini to build a RAG based in DT via LM Studio and C# (*)… A somewhat of paneled stuff that will replace my current Python bunch of scripts to do things with and without DT…
As a sample, now I get a myself recorded podcast audio, drop in a folder, rename it and then run a Python script over it. Removes noise, equalizes, transcribe, summarize, generates a blog post, uploads into my rfog.es, then puts the transcription, the blog and the summary into DT, that takes them and move to the right folders. I could drop the audio, get a window asking for the final file name, and let the panel do all the stuff.
My next step is building the RAG importing data from DT and then use thar vectorized database to query it via local LLMs… I think a MCP server into DT will ease some steps.
Another in-course project is to replace Inoreader. Take OMPL, get the news, download them, mix similar news and build a comprehensive mix, etc, and put into DT. Perhaps here the DT MCP server could ease the things (for example, getting the RSS from DT itself, sanitized).
(*) I hate Python. Really hate it. From a developer point of view, it has a lot of shortcomings and tipi-tink-stinky stuff I don’t like.
I use dvcrn/mcp-server-DEVONthink MCP server with Claude cowork - predominantly for auto filing and searching. Any new docs I put in the inbox and every morning at 07:00 Claude sorts them and places them in the correct folders. When I initially set it up Claude completely revamped my filing folders etc. Game changer..
It would be great if DT MCP Server would be offered for all versions of DT. Personally I have no use for built in DT AI integration but would love to use official DT MCP Server (instead of vibe coded OSS) with Gemini CLI and Claude as an official „add-on".
Do you mean all editions?
Yes , wrong choice of wording on my side. Would be great if you offered MCP in DT4 Standard edition in addition to Pro. If MCP is the new API, API could be made available to all users of the product.
I’ve actually done an MCP server for DEVONthink - I use DTP mainly as my email archive, and sometimes need to search something in DTP. Check it out here: