[Release] DEVONthink MCP Server

DEVONthink Artificial Intelligence
1

Hi all, I created a little MCP server for managing DEVONthink through AI agents. I know this is similar to what DEVONthink can already do through AI chat, but it’s nice to have an agent be able to interact with DEVONthink, use other MCP to do other actions and feed the results back into DEVONthink again

Screenshot
Screenshot2328×2330 393 KB

You can get the server from GitHub - dvcrn/mcp-server-devonthink

{
  "mcpServers": {
    "devonthink": {
      "command": "npx",
      "args": ["-y", "mcp-server-devonthink"]
    }
  }
}

It’s still very experimental and early, but the initial set of tools seem to be working quite nicely already! At least I couldn’t find anything exploding, but use at own risk.

Core Operations:

  • devonthink:is_running - Check if DEVONthink is running
  • devonthink:get_open_databases - List all open databases

Record Management:

  • devonthink:create_record - Create new records (groups, markdown, bookmarks, etc.)
  • devonthink:delete_record - Delete records permanently
  • devonthink:move_record - Move records between groups
  • devonthink:rename_record - Rename records
  • devonthink:get_record_properties - Get detailed metadata for records
  • devonthink:get_record_content - Read record content

Search & Discovery:

  • devonthink:search - General text-based search with comparison options
  • devonthink:lookup_record - Find records by specific attributes (filename, URL, tags, etc.)
  • devonthink:list_group_content - List contents of specific groups/folders
  • devonthink:current_database - Get the currently selected database
  • devonthink:selected_records - Lists all records currently selected in DEVONthink

Web Import:

  • devonthink:create_from_url - Create records from web URLs (formatted notes, markdown, PDFs, web documents)

Tagging:

  • devonthink:add_tags - Add tags to records
  • devonthink:remove_tags - Remove tags from records

Happy for any feedback of what else to add or what’s not working yet :slight_smile:

10 Likes
2

What I noticed (and that does not imply criticism, but you asked for feedback):

  • Why is all the code written to run asynchronously? I know that Promises are very hip, but what exactly is supposed to be happening while the promises are waiting to be fulfilled or not?
  • What is the purpose of this exercise? It appears that it provides a (fairly contrived) wrapper around the JXA functions of DT. Which everybody can call directly from a JXA script, anyway.
  • applescript/execute.ts is a wrapper for osascript -l JavaScript – so the directory might better be named javascript
  • In this file, replace(/'/g,"''") might be written more clearly as replaceAll("'","''") – you do not need a regular expression to replace a single character.
  • Same file: The error message refers to AppleScript, which is incorrect here
  • Finally, I’m not sure if result will contain a JSON string. If I understand the code correctly, stdout should just contain the output of the osascript run, which in general is not JSON. I think.
  • addTagsTo: I don’t see why one needs a database in AddTagsSchema. If one adds tags to a record, the record’s database a) is irrelevant and b) can be inferred from the record.
  • The code to set the tags is overly complicated: record.tags = record.tags().concat(additionalTags) should suffice. Using a Set seems compelling because it weeds out duplicates, but I’m confident that DT takes care of that anyway.
  • “classify.ts”: The line targetDatabase = databases.find(db => db.name() === "${databaseName}"); looks overly complicated to me. I’d suggest a try block
    try { targetDatabase = app.databases[${databaseName}](); } catch (e) { throw new Error('…') }
  • The same file uses a different semantic than DT’s classify method. In DT, a missing database parameter means “all databases”, not “current database”. I’d suggest to stick with that.
  • This check if (targetDatabase) { seems a bit pointless to me: Either targetDatabase hasn’t been set by the user, than you set it to currentDatabase. Or the user set it anyway. So the condition is always true.
  • This ${comparison ? `classifyOptions.comparison = "${comparison}";` : ""} ${tags ? `classifyOptions.tags = ${tags};` : ""} is plain ugly (sorry). Use if statements to set the classifyOptions. Ternary operators here take too long to understand, imo. Not to mention that setting an option to an empty string is not the same as not passing it at all – the latter triggers the default, the former not so much.
  • I don’t get this line at all: type: proposal.recordType ? proposal.recordType() : "group" Did you ever encounter a record without a recordType?

I stop here with just one question: Did you write all that code or did a AI spew it out?

2 Likes
3
  • Why is all the code written to run asynchronously? I know that Promises are very hip, but what exactly is supposed to be happening while the promises are waiting to be fulfilled or not?

Since when are promises “hip”?

In the end it doesn’t really matter, the MCP runs on stdin and AppleScript is most of the times blocking anyway (yes yes apps have an internal eventqueue etc but it’s still synchronous)

Agents also can’t execute multiple tools at once, and if you have multiple agents that interact with it at the same time, they would all have their own instance due to it running on stdin

  • What is the purpose of this exercise? It appears that it provides a (fairly contrived) wrapper around the JXA functions of DT. Which everybody can call directly from a JXA script, anyway.

I can’t tell if this is a serious question? Are you asking why to expose the JXA interface through MCP to an agent to interact with Devonthink?

What would be the alternative? Have an agent write JavaScript and then execute osascript on the terminal?

That would be error ridden and allow arbitrary code execution for once. The agent would also need to know the entirety of the AppleScript interface of Devonthink at all time which would overload its context

With MCP you turn any agent with MCP support into a smart AI assistance that can search, manage, organize your Devonthink databases, while also interact with other MCP servers for other things. Eg read your emails, find matching documents in DT, organize your files into groups, compose replies based on DT context, etc

Thanks for the code review! I think it would be better on GitHub :slightly_smiling_face:

1 Like
4

To abbreviate: I’m sure there are people for whom this project is useful. I’m not one of them.

Asking for feedback here and then saying that you want it elsewhere is a bit confusing, imo. As is seeing you dodging the question about the author of the code.

1 Like
5

Okay let me ask this back - have you ever used MCP? If you don’t see the value of things like this I’d guess that you didnt yet

In that case I would recommend to just try before making judgement. See if any of your ai apps support MCP connections, Claude desktop does

MCP is huge right now and allows you to hook into all kinds of things (apps, services, etc) through agents to combine those intelligently. The purpose of this project is the same as the in-app chat agent within DT4, so if you find that useful, then this will be like that, just with more functionality and more connections to other things

It’s like having a smart personal assistance. You could ask the agent to organize your entire inbox and it can rename your files intelligently and move them into the correct places

IMG_5182
IMG_51821179×2556 271 KB

IMG_5183
IMG_51831179×2556 307 KB

2 Likes
6

I have no use for AI at all. But thanks for explaining.

1 Like
7

Okay then I appreciate the critical and cynical comments but if this is clearly not relevant to you, then maybe don’t comment on it as you’re not the target audience

You comment on all of my AI tool releases on this forum with cynicism and it’s not nice to always have to justify everything.

I fully get it they it’s not for everyone, but this release is very specific to people using AI agents that support MCP, or people that already find the DT4 chat useful

Other tools that I posted in the past, like renaming files through chat queries, or classifying files through an AI agent, have been integrated into DT4 as first party tools, so there clearly is a demand. Just the existence of the chat tool within DT4 should be proof enough that some people find this of value.

4 Likes
8

You explicitly asked for comments. I gave them. Nothing cynical about pointing out possibilities to improve code, imo.

3 Likes
9

This is neat; thanks. How smartly have you found Claude using the commands? E.g., does it ever check classification results or searches to get more context from your database when it’s analyzing a document?

10

I’m interested in MCP and agents but haven’t tried them yet. From someone like yourself who has, and has been thinking about them for a while, it would be interesting to see some more of your own use-case examples and how you approach a problem to solve with them. Thanks for sharing.

11

Before making this open source, I’ve started using Claude specifically (extended thinking mode) to manage my database. I’ve been using this for a few weeks with very great success, and it’s nice to have it included in my Claude subscription without paying for extra API usage on top. I also tried it with Gemini 2.5 Pro because it has thinking baked in, but I found Claude 4 Sonnet to be more accurate most of the time.

So stuff like asking the agent to use “classify” to get potential matches and give a verdict whether it’s the correct group to move things to or not.

If you use “filesystem” MCP in combination, the agent can also upload visuals of the records if you ask it to.

I’m using a strong system prompt that explicitly explains my naming scheme and where to move things to and find it very accurate (That is Sonnet 4 with extended thinking enabled) because I also use the built-in DT4 AI features “classify” and “compare” to leverage DT4 index.

This is my current system prompt, then sub-prompt to specify naming schemes across my database, and instruct the model with commands to find groups within DT4 through tool use. It’s generated with Claude itself and whenever it makes new discoveries, I update the system prompt to reflect it.

—-

Act as a secretary and file organizer for DEVONthink document management. Your primary job is to read, update, and manage DEVONthink databases through the available devonthink tools.

File Organization Standards

Naming Convention:

  • General format: YYYY-MM Description (e.g., “2025-07 Germany Trip Documents”)
  • Exact date format: YYYY-MM-DD Description (when specific date is relevant, e.g., “2025-07-18 Concert Ticket”)
  • Maintain original language for official documents

Core Operating Principles

Safety First:

  • NEVER make assumptions about document placement or organization
  • ALWAYS ask for clarification when uncertain about categorization
  • NEVER move or delete items without explicit permission
  • ASK for permission before performing any destructive operations

Discovery and Navigation:

  • Use devonthink:get_open_databases to identify available databases
  • Use devonthink:list_group_content to explore folder structures
  • Use devonthink:search to locate documents and folders
  • Always verify current structure before making organizational decisions

Powerful AI Tools

Classification:

  • Use devonthink:classify to get AI-powered destination suggestions for documents
  • This tool analyzes document content and suggests appropriate folders based on existing organization patterns
  • When in doubt, always use this tool to understand where similar documents could be in this database. This is a very powerful discovery tool and should be used often

Comparison:

  • Use devonthink:compare to find similar documents and identify potential duplicates
  • Essential for maintaining clean organization and avoiding redundancy
  • When in doubt, always use this tool to understand where similar documents could be in this database. This is a very powerful discovery tool and should be used often

Document Management:

  • Use devonthink:get_record_properties for detailed document information
  • Use devonthink:get_record_content to read document contents when needed

Workflow

  1. Always start by understanding the current database structure
  2. Use AI classification tools to guide organization decisions
  3. Verify suggestions against established folder hierarchy
  4. Seek explicit permission before moving or organizing documents
  5. Maintain consistent naming conventions and folder structure

Your role is to be helpful, thorough, and extremely cautious with document management operations.

8 Likes
12

:joy: Hilarious interaction

2 Likes
13

I don’t trust AI so much so I would change delete to a folder called delete or something in case ai gets funny ideas.

14

All tools are approval only. Any agent will prompt you the first time you use it if you allow the tool execution

Eg “Claude wants to use the tool “remove_record” with the input “somefile.pdf”, do you want to allow this

  • yes (allow once)
  • No
  • Always allow (don’t ask again)

It’s already made with exactly this mindset - don’t trust blind execution

15

This is great - I’ve been considering writing this as well. Thanks!

16

By the way, the REAL fun happens when you plug this MCP server into Claude Code and then use claude-flow to spin up an agent swarm. Now your DEVONthink data can be analyzed at scale using as much Anthropic spend as you want to throw at it.

It would be great if DT would run with this…

17

Incorporated most points from your code review:

  • Error message renamed to JXA to make it more clear what’s happening
  • addTags - removed unnecessary ‘database’ property. Thought it was better to keep it consistent across all tools but agree that it wasn’t required
  • Simplified tag concat - I thought DEVONthink would duplicate the tags but tested this with the simpler logic and you’re right, it handles it correctly
  • classify/compared - simplified array lookup
  • simplified the complicated statements you mentioned

I’ve also added 2 more tools:

  • current_database() - returns the currently selected database
  • selected_records() - returns the currently selected records

Have you tried this? What kind of analysis could you do across your db?

3 Likes
18

Okay I was curious and wanted to try this. I explicitly told Claude Code to use multiple agents to analyze my database and it works really well. It created 3 sub-agents and used them to poke different folders and databases at the same time. Of course I only gave it permission for read tools

Screenshot
Screenshot2558×2330 495 KB

I think Gemini CLI would be perfect for this because of it’s huge context window (something like 1m tokens if I remember correct?). So it could gobble up many files, read their content and then provide a verdict on their data.

Also Gemini CLI has a huge free tier, so if anyone is curious about playing with MCP, give Gemini or Claude Desktop a try! (If you’re not scared of using the terminal) GitHub - google-gemini/gemini-cli: An open-source AI agent that brings the power of Gemini directly into your terminal.

2 Likes
19

People should also read the terms and agreements when using a free tier.

2 Likes
20

Oh yes, absolutely!

Add this to your ~/.gemini/settings.json to disable all telemetry and disallow usage for training the model. Only applies to the individual plans, upgrading to Standard through Google Cloud is excluded from usage for training purposes

cat ~/.gemini/settings.json
{
  "usageStatisticsEnabled": false,
  "telemetry": {
    "enabled": false,
    "target": "local",
    "otlpEndpoint": "http://localhost:16686",
    "logPrompts": false
  },
  "mcpServers": {
          "devonthink": {
          "command": "npx",
          "args": ["-y", "mcp-server-devonthink"],
          "type": "stdio"
        },
  }
}
2 Likes