Claude · Setup

Build an Obsidian Second Brain That Claude Code Reads and Feeds on Its Own

Four folders, one MCP command, zero plugins. The complete setup to turn your Obsidian notes into working memory that Claude consults before every project.

QQuentin Megevand
August 23, 2026 · 6 min read

A second brain that sleeps is a graveyard of notes. You capture, you file, you tag, and six months later you ask Claude a question whose answer is already written somewhere in your vault.

The problem was never note-taking. It is that nothing reads the notes.

A second brain only counts if something consults it before working. That is exactly what Claude Code does.

This guide builds the full circuit: an Obsidian vault structured to be machine-readable, connected to Claude Code through MCP in one command, with no plugins, and the prompts that keep the whole thing alive. Thirty minutes, zero dollars.

Claude AI Lab

The Claude AI Lab is my Skool community where I share my Claude systems and the more advanced modules. Access is $67/month.

Join the Lab →
What you need
1
Obsidian. Free at obsidian.md, including for commercial use. You only pay for multi-device sync (Sync, $4 per month) or web publishing, and neither is needed here.
2
Claude Code installed. This setup is verified on version 2.1.238.
3
Node.js. The connection command runs through npx, which ships with any standard Node install.
1

The vault: four folders, no more

🗂️ obsidian.md

Create a fresh vault, then four folders at the root:

/projects     one file per active project: goal, status, next step
/decisions    one decision per file: the choice, the date, the why
/context      what rarely changes: your offer, clients, stack, constraints
/resources    everything else: articles, excerpts, ideas, references

The logic behind the split is simple: each folder maps to a question you will ask Claude. "Where does project X stand" reads /projects. "What did we decide about Y" reads /decisions. "Write this in line with my positioning" reads /context.

If you know PARA, Tiago Forte's method (Projects, Areas, Resources, Archives), you will recognize the family resemblance. The difference is deliberate: /decisions and /context replace Areas and Archives, because those are the two folders an AI exploits best. A dated decision with its reasoning is the most valuable material you can hand an assistant, and stable context stops it from asking you the same questions every session.

The golden rule
Four folders are enough. Every extra folder is one more filing decision at every capture, and filing is what kills second brains.
2

The MCP connection: one command, zero plugins

terminal

One command, swapping in the path to your vault:

claude mcp add obsidian-mcp -s user -- npx -y obsidian-mcp /path/to/your/vault

Restart Claude Code. Your vault is now searchable from the terminal: Claude can read, search, write and navigate your notes.

Three things worth knowing about this connection, verified on version 2.0.1 of the package:

The obsidian-mcp server works directly on the files. It needs no Obsidian plugin, and Obsidian does not even need to be running: your vault is just a folder of markdown files, which is precisely what makes it readable by any tool.

The -s user flag registers the server for all your projects, not just the current folder. That is what you want for a second brain: it should be available everywhere.

To check the connection holds, type /mcp in Claude Code: the server should appear in the list with its tools.

Why not a plugin
A second family of connections exists, through Obsidian's Local REST API plugin. It requires Obsidian to be running at all times. The filesystem route is simpler, more robust, and enough until you need to drive Obsidian's interface itself.
3

Feeding it: what goes in, and in what shape

📥 your vault

A second brain feeds on writing little, but writing right. Three rules make your notes exploitable by Claude:

One decision per file, with the date and the why. Not a journal, a record. "2026-08-23, chose Stripe over Paddle, because X and Y." It is the format an AI retrieves and cites best, and also the one that serves you when you revisit it in six months.

Absolute dates, never relative ones. "Last week" means nothing by the time the note is reread. Write the date.

Properties at the top of the file. Obsidian's properties block (status, linked project, tags) is YAML that Claude parses effortlessly. Two or three fields per note are enough; what matters is that they are the same everywhere.

And the single most profitable habit in the system: at the end of a work session, ask Claude to write the note itself. "Summarize what we just decided and write it to /decisions with the date." The brain feeds itself at the exit of every session, with no extra discipline from you.

4

Putting it to work: the prompts that consult it

🧠 Claude Code

The circuit only pays if Claude reads the vault before acting. Three prompts cover the essentials:

Before starting any work:

Before we start, read everything in my vault related to [PROJECT]:
the project file in /projects, the linked decisions in /decisions,
and the relevant context in /context. Summarize what you take away
in five lines, then ask me whatever questions remain open.

To retrieve a past call:

Search /decisions for anything touching [TOPIC]. Give me each decision
with its date and its reasoning, newest first, and flag it if two
decisions contradict each other.

The weekly review, the one that maintains the system:

Go through /projects and tell me for each project: is the written next
step still the right one, when was it last updated, and which ones have
not moved in two weeks. Propose what to archive, update or split.
The reflex that changes everything
The question "have I already thought about this" is no longer yours to answer, it is the vault's. If you captured your decisions, the answer exists, and Claude finds it in seconds.
5

Going further: Bases and variants

🔭 obsidian.md

Two natural extensions once the base is running:

Obsidian Bases. Since version 1.9.0, Obsidian can turn a set of notes into a database view: tables, filters on properties, formulas. For an AI-driven second brain this is the ideal structure layer: a table of your projects with status and priority is readable at a glance by you, and parseable at a glance by Claude.

The same vault from Claude Desktop. The MCP server you registered is not tied to the terminal. The same mechanism plugs into Claude Desktop, and your second brain becomes available from an ordinary conversation, not just a work session.

What it changes day to day

🌅
Session start
Claude reads the project, the decisions and the context before the first line. You stop re-explaining.
🌙
Session end
Claude writes the decision note itself. The brain feeds with no discipline required.
🔁
Every week
The project review takes one prompt, run against the real state of the vault.
💶
Total cost
Zero. Obsidian is free, the MCP server is an npm package, Claude Code does the rest.

Start small: the vault, the command, and one single habit, the decision note at the end of each session. The rest of the system builds itself, because Claude is the one writing it.

The one fatal mistake
Building the perfect structure before capturing the first decision. A four-folder vault with ten decisions in it beats a complete, empty PARA system every day of the week.

Want to go further?

And day-to-day, I post one reel a day on Instagram: @quentin_iamarketing