I have only use the free versions of Claude, ChatGPT, Gemini, and Copilot. The fact that I haven’t needed to pay anything yet shows my minimal use. While much of the discussion about MCP Servers is well above my head right now, I keep hearing Davis Sparks raving about it and is very excited to use Claude on the DEVONthink MCP Server.
I have friends who pay between $20 and $100 per month for OpenAI and Claude, but are mostly working on programing and graphic uses. As generic of a question as this is, can anyone give me any idea of what kind of monthly cost I may expect to pay for Claude when using the DEVONthink MCP?
I would just like to have access to my fairly large DT Databases and find things using simple descriptions of what I am looking for. I know that this is likely a way too generic question, but since people are so nice and helpful on this forum, I thought that I would try.
If you access it via Claude Cowork/Dispatch then usage is included in a $20/month Claude Pro plan [or higher Claude Max plans] - you pay nothing more unless you go over the usage limits of your plan.
Thank you both for your replies. This does help. I now know that I at least will have to sign up for the $20/month Claude Pro Plan. I would have to see if my own use would hit the cap or not.
If anyone else has any upcoming experience of token usage once they start using the DEVONthink MCP servers, I would love to hear about it. It is possible that I will be the test case in the near future and I will document how it goes. Thanks again…
Also of note- my experience has been that flat rate use of Claude Cowork/Dispatch is considerably less expensive than using Claude Sonnet or even Claude Opus via API. When I have hit limits on Cowork/Dispatch it was usually after a volume of work that would cost me at least 5 times as much via API.
The more people use flat rate subscriptions as much as possible, the sooner the prices will converge. Sooner or later these companies have to make money instead of burning VC at an incredible rate.
A simple solution is also to use an AI provider that allows you to choose the LLM you want to use. We can thus put monthly limits on it, and what is not consumed during the month is carried over to the following month.
I’ve been using Claude Pro for a while now — in Germany for €17.85 per month with annual billing — with the MCP server from DEVONthink Pro, and also extensively in both professional and personal contexts. So far I’ve been using Sonnet 5 and I’m getting along with it very well. Even extensive research tasks, document comparisons, having documents prepared, and creating bilingual documents haven’t come close to using up my Pro data volume yet. A clear recommendation from me.
I use Deepseek v4 Pro and Flash via Opencode Go. They’re unbeatably affordable. All the models offered are hosted in the U.S. The providers follow a zero-retention policy and do not use the data to train models, with a few transparently stated “free-use” exceptions.
I haven’t yet hit any limits with text-based tasks, even with Flash, which has an extremely high cache hit rate. With OpenCode Go, you can get a lot done for €10/month.
However, I use HermesAgent as my backend, which communicates with the MCP server from DEVONthink.
@dargelm Thanks for the pointer to Opencode Go. Some digging shows that it can be used as advertised, for coding workflows, but also for chat-based research. Is your positive experience predominantly with coding, or have you used it also for broader research?
My primary use is network and server administration. I have my HermesAgent handle routine tasks.
To do this, I rely on a comprehensive set of rules that first requires an analysis, proposes a conceptual solution, requires my explicit approval for any administrative intervention, creates a backup in case a rollback is needed, implements the solution, and then uses smoke tests to verify the implementation.
Finally, complete and detailed documentation is automatically created in DokuWiki. All of this saves me a tremendous amount of time and typing work.
I benefit from the fact that HermesAgent selects the most suitable and cost-effective LLMs from a set of options, so I don’t have to use a sledgehammer to crack a nut just because I’m unnecessarily locked into a subscription. Furthermore, HermesAgent is constantly improving because with each task, it better understands its operational environment, learns from experience, and draws on its growing knowledge—so it doesn’t have to rely on the LLMs every single time.
For the initial development of solutions involving complex analyses, research, and concepts, more powerful LLMs such as Deepseek v4 Pro or Kimi3 are used. However, HermesAgent still independently uses more affordable or better-suited LLMs for subagent tasks.