Original video in Chinese.
Key Takeaway
- I built an Obsidian MCP that provides functions such as searching, reading, creating, and moving notes, as well as managing folders.
- This MCP supports multiple installation methods, including Claude Desktop’s DXT approach, remote NPM package installation, and local installation.
- Using this MCP lets you get rid of the limitations of Obsidian’s built-in AI plugins. Instead, you can directly access and manage your notes with familiar AI clients like Claude and ChatGPT.
- Combined with tools like Prompt House, you can achieve smarter note management and AI invocation.
- I emphasize the importance of context in creating value with AI, and I believe the opportunity lies in providing AI with sufficient, appropriate context.
Full Content
As a heavy Obsidian user, I built an MCP for this software. Chances are, it’s the most comprehensive Obsidian MCP server on the market right now.
In terms of functionality, it can search the vault, it can read, create, and move notes, and it can also manage the folders inside.
In terms of installation methods, it supports three:
If you use Claude Desktop, then I strongly recommend using the DXT method. This is the latest method launched by Anthropic. You only need to download the DXT file locally and double-click it, and the installation will be completed.
Or, you can use remote installation. I’ve already published the NPM package. So you only need to fill these in in the configuration file, and that’s it.
If you’re hands-on, you can also use local installation. You can use npm install to download the whole package locally, or download the whole repository and install it that way, or use Docker as well.
So why do you need this MCP?
Because with it, you no longer need to use the AI plugins inside Obsidian—honestly, those plugins are always a little underwhelming. You can keep using the AI client you’re familiar with, like the world’s best AI clients, Claude and ChatGPT, and then connect the client to your note vault so AI can access all your notes and help you manage them.
Going one step further, if you can combine it with Prompt House that I developed and set retrieval rules and so on in the prompt. When a specific situation comes up, AI first uses the Prompt House MCP to get a specific prompt, then automatically goes into your vault to look for specific notes, one step after another, without needing you to give complicated commands.
At the moment, I’ve only developed the basic functions of Obsidian MCP. So at this stage, it overlaps quite a bit with File System MCP. Because a note vault is essentially a local folder, and a note is essentially a Markdown file. Next, I’ll develop advanced features, such as automatically adding backlink functionality, and so on. You can look forward to it.
So how do you install this MCP? I’ll demonstrate the first three installation methods, because they’re the simplest and suitable for everyone.
Before we begin, you need to prepare two things.
First, you need to know the address of your note vault. For example, I created a vault on the desktop for demonstration, and its address is this string.
Second, you need to install the Local REST API plugin. Just search for it in the Obsidian community plugins. Then install and enable it. In the settings, turn on the unencrypted option, because we’ll be using port 27123 later. This API Key will also be needed later.
Let’s look at the DXT installation method first.
Go to the GitHub page and download the DXT file. You can think of it as a fully packaged installation file. After double-clicking it, it will automatically open Claude Desktop for you. Then click Install, fill in the vault address and API Key you prepared earlier, start this MCP server, and everything is done.
Next is the NPM package installation method.
The client I’m using here is Cherry Studio. Create a new MCP server, choose stdio as the type, fill in npx for the command, and @huangyihe/obsidian-mcp for the arguments—that’s the name of the npm package I published. Fill in the environment variables like this, then save, and that’s it.
Finally, the local installation method.
First, run this line of command in the terminal. Then open the client; I’m using ChatWise here. As before, create a new MCP server, choose stdio as the type, fill in obsidian-mcp for the command, and keep these two lines for the environment variables. Finally, click “View Tools.” If these tools pop up, it means the installation was successful.
For these three methods just now, one of them will definitely work for you. As for the results, I’ll show you in Claude.
I asked Claude: what notes does Obsidian have? This was to check its retrieval ability. You can see that it called the tool and quickly listed them out.
Let’s make it a bit harder: help me create a folder named 123 in the note vault, and move the Welcome note into it. Then create a note named MCP inside the 123 folder, with content that is an introduction to the Model Context Protocol.
You can see that Claude 4’s tool-calling ability is indeed the strongest at the moment, and it completed the task easily and accurately.
Honestly, among so many clients, the one I currently feel most comfortable using is still Claude Desktop. Especially when it comes to MCP invocation, its experience and results are the best.
Lastly, one more thing: why did I develop this MCP, and the Prompt House I made before?
As I said in the community: Context matters!
The model’s capabilities are already strong enough. Going forward, to let the model create value, you need to give it sufficient, appropriate context. For those of us who are not model researchers—developers and entrepreneurs—the opportunity is hidden in context.
OK, that’s the content for this issue. If you want to talk AI, want to become an individual of extraordinary capability, and want to find like-minded people, come to our newtype community. See you next issue!