Agentic Obsidian
I let Claude read and write my vault. Here’s what it did.
Last month I wrote about building a medication tracker for my wife’s surgery recovery using Claude. The point was simple: AI is changing the default from “find an app” to “build one in an afternoon.” What I didn’t say in Disposable Software is that the same shift is happening to the vault itself.
I’ve been running an AI agent against my Obsidian vault for a few months. Not a chatbot that answers questions about my notes. An agent that opens files, reads them, rewrites them, and creates new ones. It runs on a schedule. It does this while I’m at work, sitting in front of the computer, or asleep.
This is not theoretical. This is my actual setup. And I want to walk through what it actually means before the think pieces about “agentic AI” turn it into something unrecognizable.
What “agentic” actually means
When most people talk about AI and their PKM, they mean asking ChatGPT to summarize a note they paste in. That’s useful. It’s also just a slightly faster search bar.
An agent is different. It has access to tools and in this case, the ability to read files, write files, run shell commands, and call external APIs. When I talk to my agent, it doesn’t just process what I paste. It can go find the thing. It can check yesterday’s journal entry, read three related files, and write a summary without me feeding it anything.
The technical setup is Claude Code (Anthropic’s CLI tool) plus Model Context Protocol, which gives the agent the ability to interact with external systems. For Obsidian, that means it treats the vault like a filesystem because that’s basically what an Obsidian vault is. If I don’t want to use a robot to edit a note, I can just go in myself to create or edit a note.
What it actually does
Let me be specific, because “AI manages my notes” sounds either amazing or horrifying depending on how you hear it.
Here’s what my agent does on a regular basis:
Generates content ideas automatically. Every Monday morning, it searches productivity communities and recent articles, pulls out ten content ideas for this newsletter, and writes a dated file into my vault. I find it there when I open Obsidian. I didn’t ask for it that day. It just happened. (This very article started as one of those automatically generated ideas.)
Transcribes and processes podcasts. My podcast workflow is fully automated. New episodes are downloaded, transcribed locally, and written into my vault with show notes and a full transcript. I open a note and the work is done. I have links to everything the hosts talked about whether they linked them in the show notes or not.
Maintains a session log. After every work session, the agent writes a summary of what happened, what was done, what changed, what’s pending. The vault has a running memory I don’t have to maintain myself. There is no way that I could keep up with everything that it’s creating. So, I let it maintain itself.
Acts on inline instructions. If I write @COG: anywhere in a note, it picks that up and executes the instruction the next time I interact with it. For example, whenever I am dropping a note in my inbox about making a meal plan, I can put @COG: research health meals for four kids and put them in a list for the week so that the kid’s meal plan is done and then I can add whatever I want on the menu and the AI does nothing to what I write. I don’t have to remember to follow up. I flag it and move on. This also helps out determining whether I just had an idea that I need to work on or if it’s something that I want the robot to just do without having to ask me.
None of this required me to write code. I set up the system in plain markdown files, basically a set of instructions that tell the agent what to do and when.
The part that actually matters
The thing people get wrong about agentic systems is treating it as an all-or-nothing question. Either AI does everything or you don’t let it touch anything. The practical reality is that you hand it tasks with clear, bounded success criteria and you keep it away from decisions that require your judgment. It’s going to make mistakes. It’s going to misinterpret a request. That’s why you need to review what it makes and ensure that the output is what you wanted.
My agent can write a draft. I edit it before it goes anywhere. It can create a summary. I decide whether it’s accurate. It can flag a note for review. I decide what to do with it. What I don’t let it do: delete anything permanently, send anything external, or make judgment calls about what matters. Those stay with me. The guardrails are not complicated. They’re just decisions you make once about what “done” looks like for a given task.
Whether this is for you right now
I want to be honest about the current state: this setup requires comfort with a terminal, some tolerance for friction during configuration, and a willingness to think in systems rather than apps.
If you’re not there yet, that’s fine. What I’d suggest is watching this space closely, because the friction is dropping fast. Also, I’m not the only person doing this already. Six months ago this was significantly harder to set up than it is now. In another six months it will probably be a few clicks.
What you can do now, regardless of technical level, is start thinking about your vault differently. Not as a container you fill and occasionally search, but as something that could participate in your work. What would you want it to do automatically? What decisions would you never hand off?
Pick one task in your vault you do manually every week. Write down exactly what done looks like. That’s the first thing worth automating.