I couldn’t find a truly complete MCP for Devonthink, so I ended up building my own. Mostly out of curiosity—I’ve been wanting to go deeper into AI tooling anyway.
It’s not 100% finished yet. Printing functions and a few Devonthink-specific capabilities are still missing, but overall it feels pretty close to “complete” for real workflows.
Hopefully bluefrog doesn’t mind me stepping into this space
What’s not included (yet)
Print settings / print support
Download image for prompt
Chat / image engine commands
(display chat dialog, fetch models/capabilities, etc.)
Proof / compatibility
Tested and working in:
Codex
Claude
Openclaw
Hermes Agent
You’ll want something stronger than Minimax 2.7 or Haiku.
DeepSeek v4 Pro handles this well.
Below there are two tests:
Example prompt:
“Devonthink has a folder called MCP Test - Scholar Workflows. The files should include links showing how they connect to each other—but only within that folder. Add the links, and create a hub note with a categorized listing.”
I think it works pretty well. I’ve tested it with several edge cases, and fixed stuff. But, give me some ‘stress tests’ I can try to see if I can improve it.. Around 5-10. I changed the example above to better show the capabilities of a stronger model, and just what it could do [yes, noticed the link bug: fixing now]
I just found strange behavior. Not sure if it’s on my end.
So, a file has links, and they show up in the links panel but they don’t show up in the columns.
But, when I create the link manually they do; as you can see in file test.
I’ve rebuilt the database but it doesn’t change…
@BLUEFROG I’m looking forward to what the DEVONthink team can build.
I don’t even know how to program, but I managed to automate things I’ve been hoping for out of an experiment I ran in a weekend. I can’t wait to see what pros can build!
damn, just built a pile of scripts that could be replaced with an mcp without too much effort. not wasted effort, good to work through what process I want to use.
Welcome @fenrick
Definitely not wasted effort IMHO.
PS: an MCP server is not a replacement for every automation. Specific ones? Sure. But there are many instances in which true automations are more appropriate as well as superior in performance and output.
I agree, but I’ve just built a knowledge operating system combining DT4, JXA, Python and a set of Claude skills - retrieving inbox items, profiling them, mapping knowledge notes, building a knowledge graph of markdown files with a mix of alias and item links - with synthesis notes along the way; managing metadata structures, some simple dashboards automatically updated inside DT via similar means; using a mix of defined rules, LLM insights and classifer to rename and store the item correctly.
it’s been a lot of fun, but with an mcp the skills would be easier, the actions faster then embedding JXA in python; it’s actually pretty fast, but still fragile.
I would 100% agree. In fact, what I’m doing is any time I keep asking ai to do something, say transform videos through ffmpeg, I just started asking it to create a script. So, you can say I’m making the ai’s life easier
The MCP isn’t a replacement for automation, but it let’s you find, and use data (create groupings, find things adjacent to what you would normally do) in ways that you, as a slow human, simply can’t do.
My mental model is this:
an office clerk at a desk. He has one use machines, copier, stapler, puncher to do things with paper. Those are automations. One thing at a time, albeit quick with dt, but still one thing at a time. for searching, the clerk has to sort through the papers looking for one piece of information, and finding if the whole paper is useful.
With ai, and mcp, it can do all these things but in super fast mode and give the clerk a ‘computed’ pile of papers.
the mcp doesn’t replace the tools—it changes how the input set is constructed. Instead of manually searching and selecting documents, it can pre-filter, rank, and group related records so the automation operates on a much more relevant working set.
Even though i made my own, one from dt will be given special attention…
