Hi all! Version 1.5.0 has been released with a bunch of small changes (a lot of them thanks to @ebowman)
(If you use npx -y mcp-server-devonthink it should automatically update on next open)
This looks pretty cool. I was thinking that perhaps at some point DT4 might include its own scripting window like VSCode that would let users just type requests and that would support them along the way of performing actions. If you’ve seen Google’s colab tool, for example, it has an agent mode that lets you drop some files and ask a question and it will build out a notebook with comments and code step by step showing you all the visualizations you requested. You can then go back to any step and ask it to make improvements.
If DT4 had a gemini cli built into the interface, it might make it easier to run scripts, commands, and other actions with simple text shortcuts. That same text-based interface on iPhone or on Web where we can transcribe text by talking would allow typical humans to ask questions about our databases and get better agent-driven answers without needing to do the multi-step process of opening databases, searching, and then summarizing.
I think it’s for people trying to reduce cognitive load like this guy:
Personally, with all the spelling errors I see in code, I wish everybody would use AI because it seems that if they can’t even spell in English, how can I trust that they can write logical and efficient methods in 15 other languages and databases.
By the way, when the user prompts for feedback, he’s expecting positive feedback and (perhaps) simple requests that he can manage. I sometimes wonder if humans need a “/memory add” command like we have in gemini cli to explain what we learn in kindergarten about human psychology so that we can have a conversation that doesn’t always dig into the weeds. For example, if I could open a thread and see that everybody is there and being supportive…. I almost wish that these forums and chatrooms would have AI built in that would let me set my preferences so that it would rewrite my version of reality and change all the sideways comments by rephrasing them so as to give me a Happy Little Clouds view of the forum instead…
Have you seen AI code, eg the one posted here? It may not contain spelling errors. But it certainly is full of hallucinated method calls, useless error checks, and overall reminds me of a person just starting out with CS 101.
You can be certain that the only languages and databases AI can write code for/in are those for which a huge corpus is “freely” available on the net. Free in the sense of “who cares for copyright.” So, AI-written C#, Python and JavaScript code might be ok. JXA, AppleScript, Cobol – not so much so (I haven’t seen Cobol yet).
This looks and works well using it with Amazon Q Dev CLI. I think DEVONthink should raise the bar and consider further improving this MCP server or creating their own that better matches their processes. Current limitations on which model hosting providers I can use (e.g., no support for AWS Bedrock/Agents) or no support for syncing files via an AWS S3 bucket are very disappointing. I think the DEVONthink team can do better when it comes to this more technical persona of their customer base.
Yep I’d love an officially supported MCP server.
Excited to know that there’s MCP server for DT. Does it support CherryStudio config?
Welcome @hyjwonder
Just a reminder, this MCP server isn’t an official, supported project.
Is there a way to make this MCP server work with DEVONthink 3? Currently it expects the application to be named “DEVONthink” vs. “DEVONthink 3”. While changing this across the .ts files to “DEVONthink 3” does make the connection work initially, any actions like getting DBs still fail. Any help and insight is appreciated.
How would an MCP server work with a version of DT that doesn’t support AI?
The same way the current MCP Server works with DT4 - via Applescript
The current MCP Server accesses AI via Claude Cowork or whatever other AI app calls the MCP server. It does not use AI integrated into DT4.
As a newbee to installing custom connectors in Claude Cowork, can you guide me how to install this specific custom MCP?
In Claude Cowork under “Add custom connector” I have tried to paste the URL ‘https://github.com/dvcrn/mcp-server-devonthink’ as the remote MCP Server URL, but it is not connecting.
When I look in the README file I don’t see any other guidance to install the MCP in Claude Cowork?
(I just want Claude Cowork to read and search in my DEVONthink files, not to rename/move etc files, as part of my skills in Claude Cowork)
The MCP server needs to be installed on your computer.
The simplest way is to give Claude Cowork the URL and ask it to install it to work with Claude Cowork on your computer.
Once it is installed Cowork has a configuration section to define what tools you want it to be able to run always, never, or only with permission.
Thank you for your quick reply. I will try it out. If installed on my computer, how does it then know, if there are made any changes to the MCP server on Github that needs to be updated on my computer?
It does not. You have to check occasionally.
Or you can set up a Scheduled task in Cowork to check for any updates to the MCP server repository
By the way - I see you mentioned a “Remote server URL”
This MCP server is local, not remote.
While there are ways to convert it to a remote (cloud based) MCP server that is pretty advanced.
Local MCP servers are only accessible via Claude Desktop - not Claude.AI in the cloud
However once you have it working on Claude Desktop, you can use “Dispatch” on your iPhone to issue prompts/commands to run on your local computer back home.
You can also use iCloud Mirroring on your laptop when away from home - that lets you use your MacBook to run Claude Dispatch.
I have DEVONthink and Claude Desktop installed on two computers, all synced with iCloud (CloudKit).
Can I have the MCP installed on both computers? Or anything I should be aware of (except remembering to have only one Claude Desktop open, when using “Dispatch”)?
That’s the key issue - only keep Dispatch open on one computer.
Frankly Dispatch is so convenient to use you may find it’s easier to only have it on one computer since DT4 already syncs between the two computers. Since Claude Desktop has configurations which are local to your computer, I have found it to be easier to just keep the MCP on my main computer and use Dispatch and iPhone Mirroring to my laptop when away from home. That’s easier than keeping the config of two Claude Desktop installations aligned.
