That framing actually helps — I think I was thinking about MCP too much as “DEVONthink context inside Claude” instead of reusable workflows / skills.
What I’m personally less interested in is full automation (or replacing built-in tools / scripting), and more in using AI as a layer for suggestions.
For example: reviewing a set of documents and suggesting tags, spotting patterns, proposing ways to organize things.
So maybe the way to think about MCP for me is less “bulk processing” and more “AI-assisted curation”. That feels like a more interesting use case than just renaming/tagging.
P.s. ofc the ideal use case would also be “chatting with your entire DB in DT”. But I’d say this is not feasible at the moment
One use case where MCP has already been genuinely useful for me is tag curation.
Example: over time I ended up with inconsistent tags such as quick-lunch and quick-lunches (plus many similar cases). It doesn’t really matter how the inconsistency was introduced — manual tagging, AI suggestions, tags on import, existing tag assignment, etc.
Using Claude via MCP, I was able to identify these inconsistencies, consolidate, and clean up tagging across 200+ documents.
Of course, I burned through the quota doing this — but what I found valuable was not the bulk execution itself. It was the semantic review and decision support. I’m not sure I would have achieved the same result as easily with smart rules or scripting, because the problem wasn’t “apply rule X” but first recognizing that the taxonomy itself had drifted.
I agree completely but would add one really important nuance which is easily overlooked but is transformational in terms of what the the MCP can do.
Claude Cowork/Dispatch is agentic AI. Devonthink chat is a chatbox. The implications are enormous. The MCP lets you use this agentic enironment with your Devonthink data.
The difference is particularly notable because Claude can so easily set up agent teams and parallel subagents simply by natural language.
Simple example - suppose I have a Devonthink group with many nested subgroups and I want to extract the key contents of all of the documents.
With Claude Cowork using the MCP I can simply write:
Go through Group X-Devonthink-ABCDE and its recursive subgroups. Make a table of each document, a clickable x-devonthink link to its title, its date/size, and a 3-bullet-point summary of its content. Then add a narrative summary of the key documents/concepts for the overall project. Save the final report as an attractively formatted HTML document in the original parent group.
Claude with the MCP will do this, it will do it well, and it will do it quickly by spawning parallel agents for the various tasks. You can tweak the result by asking Claude “Do that again but change XYZ.” When you get it as you like it you can ask Claude to make it a SKILL so you can just say “Run my group summary skill on X-Devonthin-ABCDE.” And to preserve the skill for the future you can ask Claude to create a private Github repository with the skill.
It is not an understatement that such a workflow is transformative. Without exaggeration I have created some workflows with Claude just based on casual conversation with moments of free time in a week which rival what I could previously achieve in a a few months with the help of a freelance developer using n8n or python. I am not certain a similar outcome is even achievable with standard Applescript/JXA scripting when there are large datasets involved.
Do you know by any chance if it’s possible to find out uuids of specific documents in DEVONthink?
a use case I have in mind:
I am working with 5-10 documents, and want Claude to either summarize or compare some contents of the documents. So I tell it: “compare these documents with uuids xyz”.
I assume the same can be achieved by telling it the names of the documents. But I guess that would make it more expensive? With UUIDs, my assumption it can directly fetch and start working.
I’m impressed (again) how intelligent these AI programs are. They can already do something that was possible with trivial scripting only since several years.
Yes it does that well. And you can give it the uuid of a group and let it find the UUIDs of all documents in the group. It can also easily do this recursively if you have subgroups.
I tried that with Hermes, remote Ollama, Qwen3-coder-next. The poor thing needed more than 2 minutes to throw 27 UUIDs at me. A five liner in JXA does that in less than a second. Not even using any UUIDs for database or group.
Perhaps my setup is bad, I don’t know. Or everything would be blasting fast if the AI wouldn’t tell me every baby step it takes.
This is a trivial task. But if I really needed these 27 UUIDs, why would I want to wait over 2 minutes for an AI to give them to me?
If you want a quick one-off response to a question then Chat (or potentially scripting) is best
If you want a detailed analysis of a few hundred documents contained inside a set of nested subgroups, the ability to make the request of CoWork or Dispatch and then let it do its job unattended and notify you in an hour when it is done is transformative.
I just wanted to say thank you for setting this up. Claude has become invaluable, and having a real MCP server for this is awesome. rkaplan has the right idea of how to use this new powerful tool.
Also - for those of you new to Claude, and LLMs in general - they need an external repository to maintain a long-term memory about a project. This could be a folder of .md files, but how much nicer to have this as a versioned .md setup in a DT database? This is better than Obsidian IMHO (what many people use for Claude’s external brain).
Again - many thanks for doing this right cgrunenberg, and understanding it for us BLUEFROG (all as per usual).
I want to express my thanks to the team for creating an official MCP.
I had hoped to report here on all the cool ways I’m using it, but TBH, I’m still developing the overall use plan and converting my Cowork workflows to incorporate DTP. I can say, however, that it’s running great and I’m finding new ways to take advantage of it all the time.