I got a monthly Claude Pro to test, and everything is working super well. Great job Devonthink on the integration!!
A follow up question - I beg your patience, I know this has been discussed all over, but I am having a senior moment - when using Claude Cowork via MCP to query Devonthink, how can I know the cost ($) of the stuff I do?
It’s a flat rate. If you hit your limit it stops working until enough time goes by to reset it.
If you want to continue using it beyond your limits, it will offer for you to pay more. That is managed under Settings/Usage and Settings/Billing. Or just wait until enough time goes by so you are below your usage cap again.
Or if you’ve got the hardware you could stand up local models and pay nothing and not send all your data to big tech douche bros to monetize. If there’s sufficient interest, I’ll even write up an article.
Prior to yesterday, I was not too excited about MCP with DEVONthink. I didn’t think I had a use case. Then suddenly, I did.
I keep daily journal, which includes task management elements. One markdown document for each month, with daily entries. But I realized that I was terrible about using that information. I rarely conducted the daily, weekly and monthly reviews that I’ve scheduled. Could AI help?
The answer seems to be YES. It’s still very early days, but here’s the one query that turned me into an MCP fanboy.
In my “Daily notes.md” document in the Notes group of the Records database, tasks are formatted as lines starting with [], or ~. List the tasks, eliminating duplicates.
It wasn’t fast but it did it. And it did more. I didn’t tell it that indicates completed tasks but it figured that out and annotated the list. I didn’t tell it that my tasks are grouped by subject area, but it figured that out and properly grouped the tasks in the list. (And I’m not talking about simple markdown headings, I use styled spans for my groups yet it still figured it out.)
So I’m convinced.
Using: M1 Max w/64 Gig, LM Studio and gemma-4-12b-it-optiq
the thing that tripped me up at first: the local route isn’t the same path as the Claude integration, so it’s easy to assume you need a paid plan for it. the MCP bridge (Claude Cowork querying DEVONthink) does run through Claude’s cloud. but DEVONthink 4’s own Chat is a separate thing, and it can point straight at a local model, so nothing leaves the Mac and there’s no per-message cost.
the setup is short:
run Ollama (or LM Studio, which has a built-in server) and pull a model
in DEVONthink’s AI / Chat settings pick the local provider and point it at the local endpoint (Ollama defaults to localhost:11434; with LM Studio you switch the server on and load a model)
run only one of the two at a time, they compete for memory otherwise
honest expectations: a local 7-8B model won’t reason like Claude, so it’s strongest for search, summarizing, and tagging across your own material rather than long reasoning chains. RAM is the real limit; 16GB of unified memory is where a genuinely useful model starts to fit comfortably, and more is better. one gotcha a few people hit is the local model not getting the full document content in Chat, which is usually a context-size issue, so a smaller doc or a larger context window helps.
for a private, offline, no-subscription setup over your own data it’s genuinely good, and it’s the direction i’d lean for anything sensitive.
or ~. List the tasks, eliminating duplicates.