Obsidian MCP servers: experiences and recommendations?

Yesterday I learned about MCP (Model Context Protocol): “an open-source protocol developed by Anthropic that enables AI systems like Claude to securely connect with various data sources”.

A quick search on mcp.so gives me 24 MCP servers targeted to Obsidian.

I’m curious to learn how this could benefit me, and how to select the right one for me.

I guess, for now, my main application would be to generate new content based on my knowledge base in Obsidian.

Would you share your experiences and recommendations please?

6 Likes

Showcased thread:

2 Likes

I watched this, but it’s totally Windows based. As a not too technical Mac person I wouldn’t know how to convert this into usable knowledge. But thanks anyway.

After watching several video’s I think this one from Paul Dickson was the easiest to follow (albeit a bit long).

Hey guys,

Thank you for the shout out :wink:

If you’re looking for a Mac based tutorial Zen van Riel video is a good one. https://youtu.be/VeTnndXyJQI?si=D7ggza0YXVytV7iD

Just under 13 minutes which is a bit shorter than mine :slight_smile:

Sidenote:
I managed to get the Obsidian MCP Servers working in Visual Studio Code + Cline as well https://www.youtube.com/watch?v=YQMaVvrhVLE&t=1327s

MAC Cline Configuration with MCP Tools installed in Obsidian

{
  "mcpServers": {
    "obsidian-mcp-tools": {
      "autoApprove": [
        "fetch",
        "get_server_info",
        "get_active_file",
        "update_active_file",
        "append_to_active_file",
        "patch_active_file",
        "delete_active_file",
        "show_file_in_obsidian",
        "search_vault",
        "search_vault_simple",
        "list_vault_files",
        "get_vault_file",
        "create_vault_file",
        "append_to_vault_file",
        "patch_vault_file",
        "delete_vault_file",
        "search_vault_smart",
        "execute_template"
      ],
      "disabled": true,
      "timeout": 60,
      "command": "/Users/yourfolder/ObsidianVaultName/obsidian/plugins/mcp-tools/bin/mcp-server", \\ Might be different for MAC
      "env": {
        "OBSIDIAN_API_KEY": "yourLocalRESTAPIkeygoeshere"
      },
      "transportType": "stdio"
    }
  }
}

Windows Cline Configuration with MCP Tools installed in Obsidian

{
  "mcpServers": {
    "obsidian-mcp-tools": {
      "autoApprove": [
        "fetch",
        "get_server_info",
        "get_active_file",
        "update_active_file",
        "append_to_active_file",
        "patch_active_file",
        "delete_active_file",
        "show_file_in_obsidian",
        "search_vault",
        "search_vault_simple",
        "list_vault_files",
        "get_vault_file",
        "create_vault_file",
        "append_to_vault_file",
        "patch_vault_file",
        "delete_vault_file",
        "search_vault_smart",
        "execute_template"
      ],
      "disabled": true,
      "timeout": 60,
      "command": "D:\\YourVaultLocation\\.obsidian\\plugins\\mcp-tools\\bin\\mcp-server.exe",
      "env": {
        "OBSIDIAN_API_KEY": "yourLocalRESTAPIkeygoeshere"
      },
      "transportType": "stdio"
    }
  }
}

You will need the Local Rest API Plugin + MCP Tools installed
The MCP Tools plugin needs to be added to config above
I use Claude 3.7 with Obsidian AI Tools in this configuration but you could use your own Claude API if you like

My experiences:

  1. Claude Desktop takes about 5-10 minutes to output per prompt which is slow.
  2. You can’t interject half way through Claude making changes
  3. VS Code MCP Tools is a little different than Claude Desktop and seems to be a bit faster.
  4. The AI can see your whole vault (privacy concerns)
  5. MCP Servers and Tools should be used carefully

The best way to answer your question is to get in and try as many Obsidian MCP Server as you can in “test” Vault(s).

I’m finding it interesting to play with these MCP Tools but as it stands it isn’t something that I have enabled permenantly in my main Vault.

Thanks so much for this. That screenshot w/the obsidian plugins was just what I couldn’t figure out to get search_vault_smart working. Great API plugin, and a great MCP functionality, now working on Cline + GPT4.1 + Obsidian as a bleeding edge “AI OS”

1 Like

Dear Paul, thank you so much for your effort and very useful videos.

Let me just share my experience with MCP so far.

I have tried MCP with Claude desktop and another host called a msty studio.

My experience is one of disappointment when it comes to summarization and tabulations. I think one of the reasons is the large size of my vault (close to 4000 notes) causing token size limitations

I feel RAG based interactions are better for summarizations tabulations etc.

MCP may be good for note creation but I don’t use AI to create my notes. My typical use case would be to search for ideas and links that I might have missed.

Your thoughts?

Hi Lakshan,

Thank you for sharing your experiences with MCP +Claude. My experience with MCP with Claude Desktop has been similar. I found it useful for basic note creation and research. It can also be good as a search tool.

Personally I don’t use it in my Main Vault but I do have another AI Research Vault as a Sandbox to play with the tools. I find waiting for Claude to finish can be annoying and asking the AI to do everything takes a lot of the fun out of using Obsidian.

Msty is really good but it works outside of Obsidian.

I feel like it’s a good tool to play with but I haven’t found a life changing reasons to use it in my day to day workflow

My typical use case is to ask the agent to summarize everything I did last week, looking up all the notes in the journal, with links and all.

I don’t like auto-generated text in my vault (is forbidden in fact), so I gave the LLM another vault to play with, where can add the summaries, indexes, analysis and so on. So I work with two vaults: My personal vault and the mirror of it, seen through the eyes of the LLM

It is pretty effective for administrative things as well.

1 Like

Hi Cristian,

I think that’s a great use case and one that I can see being really useful. Instead of relying on data view queries or MOCs to reflect you just ask Claude + MCP Tools.

Have you found the results to be helpful?

The results are useful until now, I think it needs a lot of experimentation.

Summarization of the Journal, or the latest updates.

other use case is to write analysis in the form of text and link summaries (something like a TOC), . In my case Is sometimes difficult to know what’s inside the vault, specially when curating the disconnected information.

I have been heavily relying on range search regex queries such as…

start[\s\S]*?end

…but these are not too effective even with boundaries like \bend\b, especially in my native language.

So, because I also do not allow AI written lines in my vault’s important parts, my number one interest for indexing my full vault with vectors was for semantic searches.
I figured what is impossible to find with regex, I’ll find with AI (I do these in Templater .js user scripts and Codescript Toolkit run .ts scripts as well: what I cannot find a good regex for, I run that part through AI now in the script), but the sad outcome was whatever tried through AI (Obsidian AI plugins and Msty app) was not very effective.
I guess it’s okay for tasks or birthday checks in a 500-2500 note vault but it doesn’t seem to cut it for complex philological stuff for a 20k vault. I also cannot let AI embedders chunk through my stuff where important parts may get ignored.

So I’ve become my own AI now and started adding special blockID level connections I call collections. It is pretty well organized with helper .ts scripts now (a demo vault is possible but judging from the interest people take in the forum in the last year or so, not likely) which scripts were made with AI and if I was to improve on working with them, I’d definitely want AI to be helping in the script or other helper scripts.
So AI to help my script (I cannot script from scratch) and I want AI functionality in the scripts I cannot simple do with programming.

The problem with all this is that while I consciously want to improve on the way Obsidian can be used, I miss out on the necessary day to day text processing no one else will do for me.
As I’ve found you’ll need to do day to day processing enough on your own to learn the hard way, let stuff sink in from forums to know what is possible and the ideas will come unexpectedly. But something tells me AI is only 10-20% and the rest is you.
Like how successful people always tell you: 10% talent, 90% hard work (practice through repetition).
But it is exactly repetition you don’t want to do in computing.

That’s another thing I don’t like about AI. I have to experiment on when or where it is more likely to get the reliability I want. In real life terms, this can take about as long as building your own AI (or you).

Something worth exploring more for sure

One thing I have noticed with AI Tools is you can have a chat open for the day and talk about 5-10 topics over the course of the day. The next day you might do the same. Going back and trying to retrieve the most meaningful knowledge can be a task in itself. Unless you captured what resonated with you at the time. Even then finding the information in your vault later requires recall skills. The real power moving forward is going to be storing context memory over long periods of time and being able to speak to the AI as if it knows you better than you know yourself.

maybe you should try infio-copilot

after struggled so many ai puligin , I think infio is the most easier way to start use ai in obsidian

2 Likes

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.