💎 How Claude + Obsidian + MCP Solved My Organizational Problems

I thought they were just a gimmick, but MCP Servers are unusually reliable — and useful — compared to most AI tools I've tried.

As I tried to convey back in 2023 when I gave my personal take on using LLMs, I try to walk the line between “early adopter” and “wary of hype.” So the first time I heard of MCP, I was a bit dismissive — it sounded like an API for AI, maybe a bit more structured but fundamentally a hype-tastic rebrand. And even now I’m not sure I was wrong. I know what an API is, and I know what a MCP server is, but only in the loosest possible way, not in a nitty gritty technical details way. I mostly don’t work with APIs (mostly, I break things), and even scripts I’ve written myself sometimes feel just this side of magic.

But, after keeping my ear to the ground and playing with it a bit, I’ve got to admit that using Claude with MCP servers is hella useful. It has helped me clear out a backlog of organizational tasks in — forgive me the cliché — just minutes a day.

---

Claude Desktop, Raycast AI, Cursor, and probably a bunch of other AI services can can utilize MCP. Anthropic seems to have leaned in to network effects of an “open protocol” that they lead.

OpenAI has chosen a different path — from what I can tell, they seem to be leaning in to “connections.” You can connect to Github and Linear, Google Drive and Dropbox, Outlook and Gmail. So far it seems to be a curated list of services. I imagine the strategy is that they think this is easier for users to set up than something “complicated” like MCP servers, which are less point-and-click and more “edit a settings file.”

But personally, I’ve found that the connections I’ve tried with o3 are getting distinctly subpar results compared to MCP-backed queries with Claude.

One big use-case of LLMs for me is “go to these 6 relevant links that I am providing for you (say, from my own website), and compile the information there on X topic (say, the role of marriage in the ancient world) into a report in one place, with an easy way to go back to check the source to double-check your work.”

Claude does this just fine, and can do similar searches on my filesystem — which is to say, in my Obsidian notes. It will even run a search across my notes to find related files I may have forgotten, by brute-force searching a bunch of word combinations that would take me longer than punching in a single command.

OpenAI can’t access the files on my computer, and doesn’t seem to have a Notion connector. It can connect to Google Docs, though, and maybe it would behoove me to move most of my notes to Google Drive so that I can use them with Gemini and ChatGPT.

But realistically I’m not going to do that, because I have 12 million words worth of notes in my Obsidian vault, which is blazing fast, exceptionally flexible, and easy to manipulate with scripts.

Scripts Claude is impressively good at writing.

---

The collection of notes in my Obsidian vault used to be exceptionally tidy. Then I got pregnant, quit my job, changed careers, had a baby, and mostly stopped writing fiction. So now it’s a bit of a mess — or at least I think so, you can judge for yourself, most of my notes are available thanks to Obsidian Publish.

“A bit of a mess,” doesn’t mean “disorganized chaos,” though — in my case, it means that there were structured systems that evolved over time but are still relatively straightforward to align. It means that I have gone back and forth between the frontmatter required for one plugin or another. I’ve changed my post template. I’ve imported nearly 10 years worth of articles from Substack — and they’re structured more like a Substack export than my article drafts.

I’ve never found a good way to bulk edit template structures across multiple files in Notion or Google Documents or Google Keep — Obsidian makes it relatively easy. For bulk search-and-replace of Obsidian files, you can use an integrated development environment like Visual Studio Code. To edit or merge your tags, you can download a plugin. For more complicated things, like moving data from a specific subsection of a daily note into the log files that are my preference to read, I’ve written my own scripts and plugins.

With a lot of help.

People in the Obsidian Discord community have been immeasurably helpful in teaching many people, myself included, how to code. I’ve been friends with hobby coders since I discovered online gaming as an adolescent, and eventually married a computer engineer. So I had the resources I needed to fumble my way through a custom theme and a custom plugin. It took time, though — and a lot of asking for favors, from people who were themselves spending time helping me on something fairly basic instead of leveraging their own skills for more complex tasks.

So I ended up with a lot of projects on the back-burner. Things like putting together a nice theme for Obsidian Publish that could accommodate the quirks of my vault and suit my personal tastes. Reorganizing the naming schema of my vault to be easier to remember. Squaring the frontmatter key:pairs across various types of files. Making sure my file paths minimized spaces and special characters. Separating out the images in my attachments folder so it would be easier to find the ones I needed to use as profile pictures. Putting all the tasks scattered throughout my notes into one place for easy access when I have downtime — which is getting more common now that my kids are getting older.

---

I set up my first MCP server for work purposes. I do Quality Assurance for Readwise, which means that when my colleague Piotr built an MCP server for our users, it was my job to test it. It means that even before we had proper documentation for the Readwise MCP server, I had to figure it out how to make it work — to make sure it did work in places other than the developer’s machine, and that a small step off the happy path wouldn’t break it.

Before I tried our MCP server — which makes it easy to search your highlights from within a LLM interface, so you don’t have to context switch if you’re doing research in a bunch of different places — I had to figure out a sample one from Anthropic. I wanted to work from all of Anthropic’s official documentation. Since I use Obsidian, and my notes are on my local filesystem, the filesystem MCP server was the natural choice.

I followed the directions, edited my settings file, restarted Claude (a critical step I forgot the first time…), started messing around, realized I wanted to let it access more files than I’d initially bounded it to, edited the settings file…

and discovered that it is a genuinely useful tool in my toolbox.

To be clear, it does not replace the default Obsidian search — which I can trust to be rock-solid comprehensive. It’s not better than the Omnisearch — which does some kind of fancy OCR indexing of things like PDFs. But it takes only seconds to run a Claude search in addition to those other methods, and for a vault as big as mine — 12 million words and counting — having more ways to find useful things is really nice, especially when I know I have a note about something in there, but can’t remember what exact word I used, or who the author of it was, and am just going on vibes.

“What are the substacks I like that are written by dads about fatherhood?”

All of my searches on Substack came back useless, and I was getting too much noise when I searched for things like “father” or “dad” directly. But when I asked Claude, it leveraged the fact that my Readwise plugin sends over a full-text copy of every article I’ve ever read and archive to organize a nice list of what it considered the newsletters I was most likely thinking of.

In seconds, I was able to send the friend I was talking to a link to

Kevin Maguire

, who writes The New Fatherhood, and

Zvi Mowshowitz

who writes about a lot of things but is a NY dad with a nice regular feature rounding up interesting information in the realm of childhood and education.

---

When I started using Obsidian, I loved how I could have one “source of truth” for a note and embed blocks or sections in other places to use it as a reference. I avoided duplicated data, and relied on links because what

Nick Milo

laid out in his linking your thinking guides made a ton of sense.

When I had highlights and annotations from a book to spin out into more targeted, more atomic notes, I went back and forth about whether it made more sense to leave the authoritative copy in the original “notes from a specific source” file or move it out into an individual claim statement for easier retrieval and better scanability. The general nature of this system is laid out in the Konik Method for Making Useful Notes, which was awarded the 2022 Obsidian October prize for best written content about Obsidian. I got a trophy and everything.

These days, though, I’ve used a script to duplicate the content. I want the book and the claim statement to show up in searches. I started doing it — thanks to a script a coworker helped me with, before I could ask LLMs instead — before I got used to supplementing the built-in search tools with LLMs.

The main advantages of Obsidian over other tools are that it is fast and that it works offline. Linking is useful but honestly secondary. And the thing about working from search results is that having one extra isn’t a disadvantage; rather, having the file name of my claim and the book show up in the search menu saves me a click. The thing about markdown files being lightweight is that having multiple copies of the same information doesn’t slow anything down.

These days, I duplicate data all the time.

I let my article drafts sit right next to the clean copies imported from Substack instead of deleting one in the name of tidiness. After I finish processing a collection of quotes and comments on an article or book, I move it into my “annotated information” folder and then let Readwise sync a brand new copy.

I optimize for ease of recording and ease of retrieval, not tidiness.

And yet, Claude has helped me get my notes tidier than they ever were.

---

I’ve always liked the Johnny Decimal system. Johny Decimal is a method for organizing a variety of material types — it's a very "whole life" system, for example what some people would call a “zettelkasten” (I prefer the English “slipbox”) is one folder of my notes “vault.” It is is designed to help you find things quickly, with more confidence, and less stress — by helping assign a unique ID to everything that gets filed. The system buckets things into areas and categories, and works extremely well for large shared filesystems in complex bureaucracies.

It also relies on numbers, which in Obsidian, means that it makes it easier to control the order in which your folders get sorted. I’ve always liked having my folders arranged in order from “least polished” to “most complete” — raw information like the highlights and annotations from an entire book go in a folder that is higher up than my published articles. Bryan Jenks has a great “seed / seedling / sunlight / evergreen” metaphor inspired by Andy Matuschak, and I like it. There’s a beautiful flow to my folder organization, but I never really liked the numbers. They’re a necessary evil born of how file systems work.

Necessary no more. Thanks to Claude, I found the words to help me avoid spaces in my folder names (which was causing complications in some of my scripts. Now my folders go ancillary → inbox → information → narrative → nonfiction → personal … instead of 30 Information → 70 Fodder → 80 Published (etc).

---

The only article I’ve ever written that went viral was my 2021 piece Yet Another Hot Take on Folders vs. Tags, which made it to the front page of Hacker News and kicked off with the observation that most software I use wasn’t created in the 2020s, and interfaces pretty poorly with a tag-based system of organization. The main thing that irritated me was when I saw people saying that folders are “old-fashioned” and therefore “pointless.” The fact that they’re old-fashioned is one of the big reasons why they’re valuable: more programs, especially at the lower levels of the file system that I prefer interacting with, support and extend their use.

These days, as I interact more and more with tools like Anthropic’s filesystem MCP — and don’t have to wait for tools like Notion to up their AI game and figure out how to stay on the cutting edge — I feel pretty vindicated. It’s amazing to be able to whip up a script to update my filenames to have the spacing I prefer. I love that when Obsidian Bases came out, Claude helped me swap my dataview key:pairs into Bases-compatible metadata… which means I was able to get Notion-esque databases working in my note-taking app in a lightweight (and script-editable!) way supported by Obsidian core, with far fewer concerns about a plugin going out of maintenance.

In ten minutes or less, I really did go through every link in my vault that broke because I performed an operation Obsidian’s index couldn’t handle, and make the necessary fixes. Good backups, Obsidian’s graph view¹, Claude’s log file, and my own small abilities at reading scripts and debugging problems² gave me confidence that the fixes were not ruining my vault.

In ten minutes or less, I really did cycle through all of my old daily notes and reformat their information into skimmable log files, because although I’ve known for years that themed logs were more useful than daily notes, I never had time to actually move my data around by hand, or learn to write the scripts that would do it for me.

It really did spin up a lovely Obsidian publish theme that matches my Substack logo.

With Claude and the shockingly easy to install³ MCP Tools plugin, I don’t even have to properly prompt Claude to write my scripts — I can just say “go through the files in this folder, and figure out the patterns, and write a script to put information like A into location B.” A few iterations later, and bam.

My longstanding organizational problems are solved, and my vault is a lot more useful than it used to be.

1. Some people think that the graph view is just a pretty gimmick, but they are wrong. The graph view is not my favorite way to navigate but it is immensely useful at providing an overview of a vault’s structure, which means you can identify sync problems or broken links at a glance.

2. Some people feel comfortable vibe coding entire apps even though they’ve never written a line of code in their lives. I am too paranoid for that sort of thing. I never would have tried this method of getting Claude to write scripts to fix my vault if I didn’t have some familiarity with how scripts work, how plugins work, how javascript works, etc. Claude is great at taking someone with level 2 coding skills to the point where they can accomplish level 4 programming tasks. You may feel differently, but I wouldn’t trust it for truly complex things or things I wasn’t experienced at debugging. Even so, I now understand why people say it can 10x their engineering speed, and if I had gotten nothing else out of this, I’d still be glad I took the time to familiarize myself with this process. It was a nice, safe project for someone at my skill level, and I learned a lot.

3. Especially compared to other AI tools that integrate with Obsidian — I never did figure out Smart Connections or, well, any of them really. MCP Tools is the first time I’ve actually gotten an AI plugin to do what I wanted it to.