If you've got an 📝Obsidian vault full of notes and you're using 📝Claude Code, the obvious question is: can Claude just read all of this? Yes. And it's easier than you'd think.

The connection runs through the 📝Model Context Protocol (MCP) — the same protocol that lets Claude Code talk to databases, APIs, and other tools. Point an MCP server at your vault, and Claude gains tools to read, search, and write to your notes mid-session. Your vault becomes context Claude can pull from without you copy-pasting anything.

Here's how to set it up, what to watch for, and how I've ended up structuring my own workflow.

The quick version

If you just want Claude to have some context from your vault, you don't need MCP at all. Drop a 📝CLAUDE.md file in your vault root with your preferences, current projects, and anything you want Claude to know about you. Open Claude Code from that directory. Done — Claude reads it automatically at session start.

This is genuinely useful. It's also manually maintained, which means it goes stale the moment you forget to update it. But as a starting point, it's hard to beat.

Setting up MCP for full vault access

For real retrieval — Claude actively searching your vault mid-conversation — you need an MCP server. Several exist:

obsidian-mcp is the simplest. It reads your vault files directly from the filesystem. Obsidian doesn't even need to be running. Install and configure in one step:

{
  "mcpServers": {
    "obsidian": {
      "command": "npx",
      "args": ["-y", "obsidian-mcp"],
      "env": {
        "OBSIDIAN_VAULT_PATH": "/path/to/your/vault"
      }
    }
  }
}

Drop that in your claude_desktop_config.json or .mcp.json and restart Claude Code. You'll know it's working when Claude can respond to "search my vault for X" or "read the note called Y."

mcp-obsidian takes a different approach — it talks to Obsidian through the Local REST API community plugin, which means Obsidian needs to be open. More moving parts, but it can access Obsidian-specific features that filesystem-only servers can't.

obsidian-mcp-tools adds semantic search and Templater integration on top of basic vault access. Worth a look if you're already using Templater heavily.

A deeper walkthrough of all four major servers and their tradeoffs is in 📝Obsidian MCP + Claude Code.

What works well

Project context on demand. The biggest immediate win is that Claude can pull your project notes, design decisions, and meeting transcripts without you hunting for them. If you're disciplined about note-taking, this is a real multiplier.

Session logging. Set up a daily notes workflow where Claude reads today's note for context and appends a summary at session end. Your daily notes become a rolling log of AI-assisted work. I used this pattern for months before building something more structured.

Code + knowledge in one session. If you're working on a codebase and your vault has architecture notes, API specs, or debugging journals, Claude can cross-reference them in real time. This is where the setup stops feeling like a gimmick and starts feeling like infrastructure.

What to watch for

Token cost at scale. Most Obsidian MCP servers load entire files into the context window. With a small vault this is fine. Past a few hundred notes, you're burning tokens on irrelevant content because the server can't distinguish between a note that matters and one that just matches a keyword.

Write access is a double-edged sword. Claude can overwrite your notes, reformat content, or create files that don't fit your structure. There's no separation between AI-authored and human-authored content. If this makes you nervous, configure read-only access — you'll lose session logging but keep your notes safe.

No access control. The MCP server exposes your entire vault. Everything — personal journals, client notes, health logs — is equally visible to any query. There's no way to scope what Claude can and can't see.

Desktop only. The MCP integration requires a local process, so it only works with Claude Code on desktop. Switch to Claude's web or mobile app and your vault context disappears.

How I structure my own setup

I ran this exact pattern for a while — Obsidian vault, MCP server, CLAUDE.md, daily notes. It worked until my vault grew past the point where keyword retrieval was useful and I needed access control over what different agents could see.

That's when I built @MythOS, which handles the retrieval and permission layer that Obsidian wasn't designed for. But I know plenty of people who still run both — Obsidian for local note-taking and drafting, MythOS as the AI memory and retrieval layer. Claude Code supports multiple MCP servers simultaneously, so the two sit side by side without friction.

If you're starting out, the Obsidian vault setup is the right first step. You'll learn what you actually need from AI memory by using it — and that's worth more than picking the "right" tool upfront.

FAQ

Do I need Obsidian installed for this to work?

It depends on the server. obsidian-mcp reads files directly from the filesystem — Obsidian doesn't need to be running or even installed. mcp-obsidian uses the Local REST API plugin, which requires Obsidian to be open.

What if I only want Claude to read my vault, not write to it?

Most MCP servers let you configure read-only access. Check the server's documentation for a read-only flag or remove write-related tools from the config. This eliminates the risk of Claude modifying your notes.

How do I know which notes Claude is pulling into context?

Ask it. Claude will tell you which tools it called and which notes it retrieved. You can also check token usage — a sudden spike usually means a large file got loaded.

Can I use this with Cursor or other AI editors?

The MCP server pattern works with any MCP-compatible client, not just Claude Code. Cursor, Windsurf, and other editors that support MCP can connect to the same server with similar configuration.

What's the best vault structure for AI retrieval?

Keep notes atomic — one concept per note. Use descriptive filenames and consistent folder structure. Avoid deeply nested hierarchies that make search results ambiguous. The same principles that make a vault useful for you make it useful for an AI agent.

📝Obsidian MCP + Claude Code — detailed comparison of the four major MCP servers and their configuration
📝How Obsidian Became the Default AI Memory Layer (and Where It Hits a Wall) — the broader pattern and its structural limits
📝Obsidian Claude Memory: What Works, What Breaks, and What's Next — companion piece on the Obsidian + Claude memory pattern
📝Obsidian — reference memo on Obsidian itself
📝How I Keep Claude Code's CLAUDE.md Lean with MythOS — my approach after outgrowing vault-based memory
📝Model Context Protocol (MCP) — the protocol that makes the connection possible