-
Had you already been using Claude today, potentially already moving you toward your limit?
-
We don’t control how efficient or frugal Claude, or any AI model, is. The MCP server provides access to DEVONthink databases and commands. AI decides what tools, if any, it wants to use and its responses, especially to naïve prompts or when processing a larger volume of information, are going to be as wordy as it wants to be. We have made many tweaks to minimize token use from our side of things. External AI is outside our control.
The entire size of the group I run the test is 212mb.
To answer the questions of Jim:
- I did use, Claude today, but only marginally and some trivial things. I actually have never exhausted the usage quota before.
- Perhaps, calling Claude from within chat would be still economical and give more control over which model is being used…the 10 bucks I have added as Claude API credit lasted me until the credit has expired
and I was doing similar things as today’s test with renaming and tagging.
Export one of your MD files too your desktop. I’m sure it will have kind “E-book” there, too. Then use the search function in the forum to fix it. Or check the default app for MD files in finder.
The token usage of all internal chat commands is optimized, either depending on the Settings > AI > Chat > Usage option or due to the known task, e.g. bulk summarizing or tagging. But MCP usage is not controlled by us.
This is all a good reminder of something people often overlook: Compute has a cost.
Asking simple questions or for simple tasks to be done comes at less of a cost than asking for more complex things, e.g., trying to handle 500 documents.
1 Like
Suggestion on speeding up MCP usage and reducing cost - If you reference the target document(s) or group(s) by X-Devonthink ID or UUID rather by name or other search criteria, it is much more efficient.
1 Like
Confirming - there is no plan to offer MCP in a standard edition of DT4 as it is gated beyond “AI” feature set? > I have 0 plans to use built-in AI assistant but would find it super useful to be able to leverage DT-approved MCP.
Thanks!
As I’ve mentioned before: today I used Claude mostly for fairly trivial tasks (search, small changes, etc.), and my quota usage there was probably around ~5%.
I’ve actually never hit the quota limit before. For comparison, I’ve also used the Claude API via DEVONthink Chat + API key in the past, and even my $10 credit ended up expiring after a year before I could fully use it 
That’s why I was a bit surprised that with MCP I reached the limit relatively quickly.
That said, working directly from the Claude app feels much more convenient to me personally, and for occasional usage or individual document changes I think this integration is fantastic 
I’m mostly sharing this as early feedback / an observation — it made me curious how MCP usage scales for larger workflows over time and whether there are best practices for keeping usage efficient.
That is correct. AI access is a Pro / Server feature.
Quick update: I’ve now tagged around 200 documents with Claude via MCP, and quota usage has only moved up by ~11% — not bad, I’d say 
A couple of questions came up while testing:
-> Would it help to start new DEVONthink-related tasks in a new Claude chat? My assumption is that this could reduce accumulated context size — and potentially lower quota usage as well.
-> Would switching to a cheaper model help keep quota usage lower? For example, using Haiku 4.5 instead of Sonnet 4.6 for large-scale tagging or changing the names of a large group of documents…?
p.s. and also want to say thank you to the DEVONthink team for such a cool feature!
Using a lower-tiered model for such tasks would certainly be cheaper.
1 Like
With the mcp function, is the ai function of DT itself not necessary? What is the advantage of using DT? Is it ultimately the same as obsidian?
Which AI function exactly? The chat assistant? See Also & Classify?
I’m not sure what you’re asking but DEVONthink and Obsidian are certainly not the same application in many respects.
Can MCP completely replace The chat assistant?
Mostly. But the chat assistant is able to better control the costs and also to e.g. highlight occurrences of words/strings in the currently viewed document (e.g. to find spelling/grammar issues or to highlight important stuff) without modifying it.
For bulk tasks (e.g. renaming, tagging or summarizing via chat) I would definitely recommend the built-in commands or smart rules as they’re both faster and more efficient than e.g. Claude Cowork.
just curious: if for the bulk tasks you’re recommending built-in features or chat for better costs control.
what are the use cases you’d recommend MCP for?
And what would be the easiest way to reference documents from within Claude? Example: DEVONthink chat ai works on selection.
how can I give Claude with MCP a better scope? Like uuids
What is the benefit of trying to make Claude.app function like the Chat assistant? The assistant already is location and selection-aware and, as mentioned, offers finer-tuned controls of models and costs. Accessing external AI in DEVONthink is a particular experience with benefits over the inherently “blind” use of a bespoke AI application.
The MCP server functions best for processes, i.e., skills. Naïve prompts like, “What databases do I have open?” are fine to see if the MCP server is connection but what would be the point of such a query in practice?
And while it’s possible to use the MCP for renaming or tagging documents, is that really the best use when those tasks can be accomplished via batch processing, smart rules, or scripting within DEVONthink? And those processes would be more performant than using the MCP server for such a task.
Early on while getting your bearings with the MCP, there will be many unoptimized things you’ll do. That’s expected. But as you try to use it as part of your routine, you’ll come to find certain things are not ideally done by MCP just as there are many things YOU can do faster and more efficiently than AI, and with fewer surprises.
3 Likes
That makes sense — and I think this is exactly where my confusion comes from.
For small tasks (rename a few files, assign a few tags), I’d personally just do it manually because it’s faster than opening Claude and prompting.
For large-scale tasks, you mentioned built-in commands, smart rules, and scripting — which also makes sense from a performance and cost perspective. And we also that MCP burns the quota fast here.
So I’m trying to understand where MCP sits in between these two extremes.
The answer is undoubtedly: it depends. There is no universal use case for MCP, just as there’s no such case for AI (or DEVONthink for that matter). They are all tools. But having a tool doesn’t mean you have a use for it. I didn’t need a soldering iron until I had some custom wiring to do. Now that the wiring is done, I still have it but don’t often use it. It’s there if I need it but I don’t try to use it to toast bread or heat my coffee 
This is my current axiom about MCP: Don’t think ‘prompt’; think ‘skills’. View MCP not as something to chat with as much as something to delegate tasks to. Tasks you need to define (and refine) and provide to the AI as skills. So you still may “have a word with AI” but it is less likely to be interactive like a chat.
Also, while the MCP is not meant to (nor will it) replace automation, it also provides a bridge there. Not everyone is inclined to write a script, etc. and the MCP could allow for some level of automation for them. This can be advantageous with conditional automation, like DEVONthink’s Chat - Continue if smart action does. This too will require some effort and tweaking to polish but it is one potentially useful aspect of MCP.
2 Likes