Skip to content

Obsidian + Claude Code as Personal Operating System

Category: concept Last updated: 2026-04-03 Status: draft

Summary

Obsidian + Claude Code is a pattern for building a personal AI operating system: the user writes everything in an Obsidian vault (markdown files with bidirectional links), and Claude Code reads the vault — including its relationship graph — to act as a context-rich, always-current agent. The vault replaces the session chat log as the source of truth; agents read and execute, humans write and own. Described in a Feb 2026 podcast by "Internet Vin" and popularized by Greg Isenberg.

Details

Core Architecture

The system has two distinct roles that must never blur:

  • Human → writes the vault (daily notes, projects, beliefs, people, meeting notes)
  • Agent (Claude Code) → reads the vault, suggests, executes — never edits directly

This separation preserves the vault as a reliable, uncorrupted source of truth. The agent session is ephemeral; the vault is permanent.

Why Obsidian CLI Changes Everything

Standard Claude Code reads file content. The Obsidian CLI gives Claude Code access to the vault's relationship graph — the bidirectional links between notes — not just their text.

This means Claude Code can: - See which projects are connected to which people - Follow chains of thought across notes written weeks apart - Understand how an idea evolved through linked daily notes - Traverse the knowledge graph, not just search it

Without the Obsidian CLI, you have a file reader. With it, you have a brain reader.

Vault Structure That Enables Agents

The vault mirrors how a brain actually works: by connection, not hierarchy. Recommended structure: - Daily notes — dated stream of consciousness; what you did, learned, decided - Projects — one note per active project; linked to people, meetings, decisions - People — one note per person you interact with; linked to meetings, projects - Beliefs — opinionated notes on how you think things work; the most durable content - Meetings — timestamped notes; linked to people and projects

When this structure is consistent, agents can navigate it without instruction.

Reference Files: Eliminating Re-Explanation

Every Claude Code session starts cold — the agent has no memory of previous sessions. The traditional response is to re-explain context at the start of each session.

Reference files eliminate this friction. A reference file is a single note that captures: - What a project is and why it matters - Current status, blockers, and next steps - Key decisions already made (and why) - Links to related notes

At session start, instead of re-explaining, you load the reference file: "read my context file for Project X." The agent is immediately productive.

The vault is the session memory. The agent session is just a working layer on top.

Custom Slash Commands

Claude Code supports custom slash commands. Internet Vin's operating system includes:

Command Function
/context Load full life + work state from vault
/today Read today's daily note and orient
/close-day Summarize the day, surface patterns, prep tomorrow
/ghost Write content in Internet Vin's voice from vault material
/challenge Challenge a current belief or decision
/emerge Surface emerging themes across recent notes
/drift Detect when behavior is drifting from stated values
/ideas Generate startup ideas from vault content
/trace See how an idea evolved across notes over time
/connect Bridge two domains the user has been circling
/graduate Promote a daily thought into a permanent asset (project, belief, entity)

These commands make the agent's relationship to the vault repeatable and predictable. Instead of figuring out how to use Claude Code each session, you invoke a command.

Inline Delegation

One of the most powerful patterns: write a single sentence in Obsidian, and the agent handles everything downstream.

Example: "Schedule a follow-up email to [Person] next Thursday re: Q3 proposal."

In a traditional system, that sentence sits in a note and you remember to act on it. With Claude Code reading the vault, it becomes a task the agent can execute autonomously — finding the email context from the People note, drafting based on meeting notes, scheduling for the right time.

Writing becomes delegation. The more you write, the more you can delegate.

Writing as Leverage

The deeper principle: writing is not just documentation, it is leverage for AI delegation.

  • More context → agent acts more accurately
  • More history → agent surfaces patterns you can't see
  • More structure → agent executes more reliably

The vault compounds. An agent reading 2 years of notes is categorically more useful than one reading 2 weeks. Unlike chat sessions (which reset), the vault accumulates. The system gets smarter as you use it.

"Markdown files are the oxygen of LLMs." — Greg Isenberg

Privacy and Trust Calibration

Giving an agent access to your second brain raises real questions: - What data are you comfortable with Anthropic seeing? - Does the vault contain information about other people who haven't consented? - What happens if vault content leaks?

Internet Vin's approach: use a separate, curated "agent-facing vault" that contains only what you're comfortable sharing, rather than granting access to every note ever written. Some sections of the vault (therapy notes, relationship details) remain agent-excluded.

Obsidian Plugins for LLM Workflows

Several Obsidian plugins are particularly valuable when using this pattern:

  • Dataview — runs queries over YAML frontmatter in notes (tags, dates, source counts). If the LLM adds structured frontmatter to wiki pages, Dataview generates live tables and lists without manual maintenance. Example: TABLE updated, status FROM "wiki/concepts" SORT updated DESC.
  • Web Clipper — browser extension that converts web articles to Markdown and saves them directly into the vault. Essential for quickly getting raw sources into raw/articles/.
  • Marp — renders Markdown files as slide decks in the Obsidian preview pane.
  • Image download hotkey — in Settings → Files and links, set attachment folder to raw/assets/. In Settings → Hotkeys, bind "Download attachments for current file" (e.g. Ctrl+Shift+D). After clipping an article, hit the hotkey to localize all images. Lets the LLM reference images directly rather than relying on external URLs.

Note: LLMs can't process Markdown with inline images in one pass — have the LLM read text first, then view images separately.

Relation to Other Patterns

This pattern is closely related to but distinct from:

  • concepts/llm-knowledge-base (Karpathy's pattern): focuses on external sources (articles, papers) compiled into a wiki. The Obsidian OS pattern focuses on personal output (daily notes, beliefs, decisions).
  • concepts/personal-knowledge-management (dSebastien's pipeline): focuses on the input side — capturing and structuring knowledge from reading. The Obsidian OS pattern focuses on the output side — using that knowledge to drive agent execution.
  • concepts/harness-engineering: the same read/execute discipline, but for software engineering. Obsidian OS applies the same philosophy to personal life and work.

Key Claims & Data Points

  • Obsidian CLI exposes relationship graph (not just file content) to Claude Code — [source: gregisenberg_transcript, gregisenberg_thread]
  • Without re-explanation overhead, the agent session starts immediately useful — [source: gregisenberg_transcript]
  • 99.99% of people won't do this because it requires reflection + setup — [source: gregisenberg_thread]
  • Once the vault exists, the agent stops being generic and starts thinking in your voice — [source: gregisenberg_thread]
  • "One sentence in Obsidian → agent handles the rest" — [source: gregisenberg_transcript]

Open Questions

  • What is the exact Obsidian CLI interface — is it an MCP server, a CLI tool, or an Obsidian plugin that exposes an API? (raised by: concepts/obsidian-claude-code-os, 2026-04-03)
  • How do reference files stay up-to-date — is maintenance a manual discipline or can agents update them? (raised by: concepts/obsidian-claude-code-os, 2026-04-03)
  • At what vault size (notes, words) does the relationship graph become too large for Claude Code to traverse effectively? (raised by: concepts/obsidian-claude-code-os, 2026-04-03)
  • How do you handle multi-person vaults or shared team vaults — does the pattern transfer? (raised by: concepts/obsidian-claude-code-os, 2026-04-03)

Sources