Sure, see above (“MCP server and its implementation”)
I love this!
I’ve been using AlterHQ for AI-related use cases, and has been a real time saver for me these past two months. I see a lot of value on the OP’s approach.
I do think the warnings in this thread are valid. However, I find that if I’m careful and back up often, I can explore what these tools can do without worry.
If you summarize the documents or analyze the contents of the documents in any way then how could the MCP server not have access to the entire documents?
Can you explain what types of documents you will be using and what types of tasks you want to accomplish with the MCP server.
If security is a primary concern then the best plan is to use a local LLM and DT4.
Thank you!
I am not a Claude user, but I tried to use Raycast to install this MCP, and it failed with the error message “sh: mcp-server-DEVONthink: command not found, code: 127.” I am not very tech-savvy. Is there anything wrong?
The MCP server can and should have access to the documents. However:
- It only indexes the data you want it to index. You can explicitly exclude data you dont want included.
- What the MCP server has access to is less relevant - it’s what you allow the AI agent to read. That’s where the privacy concern becomes a real issue. And you control tha access using permissions on the MCP server.
The MCP server serves as the control layer between the AI agent and DT (and other data sources).
Please see my response to rmschne above for examples.
Yes tried - too slow on the Mac Air M4, and also doesn’t allow for cross indexing with other data sources.
Why do you believe that the author of the MCPa server is more reliable at protecting your security than Devontech, who have a decades-long commitment to a fanatical security philosophy?
DT4 clearly only sends to AI the document(s) or group(s) which you select.
Fanatical? ![]()
And which are not excluded from the chat, see Info > Generic inspector.
/na/ntas/ ?
I think that is a fair statement.
But meant with extreme admiration.
You may have missed it. I’m building my own. Hence excited that OP did the same.
OK that is cool.
What functions do you anticipate your personal MCP server doing that you could not do with Applescript and/or Keyboard Maestro?
I get the idea, but seems more of a wish than a proven thing. I’m old fashioned, I guess.
I put such things into my “reminder to-do” [Note 1] system, which links back to the doc in DEVONthink–if a useful added task but mostly not worth my time). I’d have to do a lot of validation and testing before relying on AI to figure out this sort of stuff to ensure as reliable as a “reminder to-do” system.
And, of course, I wouldn’t want that sort of information somewhere outside my control.
Note 1: My “reminder ToDo system” evolution. Daytimer paper → Lotus Agenda → Ecco Pro → Outlook → OmniFocus → Things → macOS and iOS Reminders
Hi all, I released 1.4.0
Comes with a few fixes that caused the npm -y mcp-server-devonthink to fail and added a new update_record_content tool to update the text content (plainText, richText, HTML) of records (was requested on github)
Been using this for a while now and very happy with it. Especially liking it for documents where I’m not really sure where to put them. The agent uses classify together with analyzing different folders to give great suggestions on what to add, modify or where to put things. It’s great at puzzling together solutions by combining DEVONthink search index, classify and other ai tools.
Here’s a real world example from earlier today. The LLM gave me a bunch of options based on the naming and organization scheme I taught it in the system prompt, then proceeded with the option that I accepted.
It looks to be executing mcp-server-devonthink as a command through sh. Try to specify the full npx path on your system. If you don’t know it, open a terminal and type which npx
If you installed node through homebrew, it’s probably somewhere in /opt/homebrew
So:
- Command: , eg
/opt/homebrew/bin/npx(checkwhich npx) - Arugments: -y mcp-server-DEVONthink
Alternatively, you can also git clone the repo, or install it globally with npm install -g mcp-server-devonthink, then just run mcp-server-devonthink. May still need the full path, so on the terminal (which mcp-server-devonthink)
Just tested this on my raycast and works
Thank you so much!
Very interested to try this out. I’d just pulled a couple of my databases out of DEVONthink and was keeping them in the plain file system, since that allowed me to interact with documents using MCP, which really is a killer feature. Will be very interested to see whether this might be replicated here.
This is pretty nice, thanks!
Thank you for creating this and sharing it! If you make major updates, do share here also. Looks very promising.
I created a PR that adds more capabilities to this, especially utilizing the new AI features.
Kudos to the Devon team for creating robust AppleScript interfaces!
Honestly - it was way easier for me to understand how the AI features in DT work by using it through MCP. Super cool stuff.
Enjoy!
PR link: AI support by ebowman · Pull Request #4 · dvcrn/mcp-server-devonthink · GitHub



