Introducing DEVONthink's MCP Server

Blog
1

mcp
mcp900×900 50.3 KB

MCP (Model Context Protocol) is a mechanism that allows AI assistants to communicate with apps to accomplish tasks for you. On macOS Sequoia and later, DEVONthink now has its own MCP server built-in, presenting new opportunities to access and work with your databases from outside the application.

First, set up DEVONthink’s MCP server in Settings > AI > MCP by enabling the server for one of the built-in providers: Anthropic Claude, OpenAI Codex, or the Hermes Agent. Then switch to that application and try describing what you want, e.g., Find all my notes about project X or Create a summary document in my Research database's inbox. The AI checks the available commands and carries out the task, including: reading and summarizing records, creating and updating documents, moving items between groups, and adding tags or annotations, etc. The more specific and thorough the prompt, the better the result.

Information provided via MCP responses, like the content of invoices you’ve requested to be examined and graphed, is sent to the AI provider’s servers online. But DEVONthink reinforces data privacy by redacting or obscuring personal information like credit card numbers, email addresses, etc. if you’ve enabled Redact sensitive content in the MCP settings.

Also, the AI acts on whichever databases are open at the time. If you don’t want the external AI to “see” an item, enable Exclude from Chat & MCP in the Generic Info inspector. For databases, enable the same option in the File > Database Properties or merely close that particular database.

13 Likes
2

So, I got LM Studio running on the same machine and I can chat with it - but how do I know if it also can use the MCP server? I sent a request to a model to find the tire size of my car based on the documents in my main database, and it successfully searched the database, gave me suggestions and links to related documents. Is this already the MCP in action, or is this a chat feature? Is there any way to see that the MCP is being used?

3

Afaik, chat is limited to documents and/or groups that you select.

4

You need to setup MCP access in LM Studio and you’ll do your chatting there, not in DEVONthink.

image
image3138×2020 406 KB

1 Like
5

Can you provide an example of how to configure L,Studios json to point to the Devonthink MCP server

6

Bear in mind, we don’t do support for third-party AI applications, but this is the stdio setup for accessing DEVONthink MCP in LM Studio on the same Mac. This also assumes you have DEVONthink installed in the root of the/Applications directory…

{
  "mcpServers": {
    "devonthink-mcp": {
      "command": "/Applications/DEVONthink.app/Contents/Library/LoginItems/DEVONthink MCP.app/Contents/MacOS/DEVONthink MCP",
      "args": [
        "--stdio"
      ]
    }
  }
}
2 Likes
Anybody have DT4.3.1 working with LM Studio 0.4.15+2?
7

Working now. Thank you. Would have guessed it involved connecting to the DEVONthink mcp server via tcp.

8

While it may be possible to use a web transport in LM Studio, it’s not the option I usually think of, so I tested with stdio.

9

I got it working, awesome! Now I just wondered why the context tokens did not reset to zero when starting a new chat, but that should be a minor problem. Thanks a ton!

10

You’re welcome.

PS: What model are you using?

12

Thank you for including an MCP server for DEVONthink. Made my day :slight_smile:

3 Likes
13

I am currently testing medium-sized local custom models where I get most parameters into the smallest file and RAM size, so I got a lot of RAM left for the context window:

huihui-qwen3.6-27b
supergemma4-31b
osmqwopus-3.6-27b-v2
… for example. I have 64 GB, and depending on context, RAM can spike up to 50 GB :sweat_smile:

14

More than most - good on ya!

15

I have a question regarding the masking of sensitive Data that is sent to the AI when using MCP. I have the option turned on. Is there a documentation what Data is masked and not send to the AI?

16

URLs, email addresses, credit card numbers, authentication tokens (e.g. API keys) and labeled secrets (e.g. “Password: …”). This might be extended in the future but e.g. phone numbers or bank account numbers are hard to distinguish from valid numerical data.

17

Thanks for the clarification — that answers my original question.

A follow-up thought: would it be possible to make the masking configurable in a future version? Two use cases I’m running into:

False positives I’d like to exclude from masking:

  • Danish phone numbers (8-digit sequences) — caught by pattern matching but not sensitive in my workflow

  • IP addresses — just context in network/log documentation

  • Order and case numbers — critical reference data that gets stripped along with actual PII

Patterns I’d like to add to masking:

  • Danish CPR number (DDMMYY-XXXX) — the Danish social security number, extremely sensitive under GDPR, used heavily in healthcare and public sector documents, and absent from generic PII libraries

I’m using the API with a Zero Data Retention agreement, so Anthropic’s side is already covered — but having granular control over what DEVONthink masks before the prompt is sent would complete the picture nicely for regulated-data workflows.

18

Could you send me some example documents via PM so that we could try to reproduce this? Phone numbers shouldn’t be stripped at all so far.

19

This has already 10x’d the value of my DEVONthink databases to me. Thank you so much.

3 Likes
20

I don’t see an option for “AI” in settings. Is this only available to Pro and Server versions? If so, fine. But I don’t see anything here saying so.

21

Yes, generative AI and MCP is only available in the Pro and Server editions.