Mythos

Obsidian MCP is the ecosystem of 📝Model Context Protocol (MCP) servers that expose an 📝Obsidian vault to AI agents, split between servers that run inside the app and servers that read the vault folder directly.

No official Obsidian MCP server exists. Obsidian's published roadmap contains no AI or MCP work, so every implementation is community-built, and they divide into two architectures with different requirements and different capabilities. In-app servers reach live application state — the active file, vault metadata, the command palette — but need Obsidian open. Filesystem servers read Markdown from disk and run headless, with no awareness of what the app is doing. The distinction determines what an agent can see and what the user has to keep running.

Key Facts

  • Category: Integration layer between a Markdown vault and MCP-compatible AI clients
  • Origin: Community-built from 2024 onward; no first-party server from Obsidian
  • Two architectures: In-app (via the Local REST API plugin) and filesystem (direct folder reads)
  • Notable implementations: Local REST API with MCP, MCPVault, obsidian-mcp, obsidian-mcp-tools, smart-connections-mcp
  • Transport: Local — stdio or loopback HTTP, typically 127.0.0.1:27124 for the in-app path
  • Access model: Whole-vault. No per-note permissions, no sandboxing, no audit log
  • Prerequisite: A machine the user runs — a desktop with the app open, or a server holding a synced vault

How It Works

In-app servers. The Local REST API community plugin runs an authenticated HTTPS server inside Obsidian and now exposes an MCP endpoint of its own, so a client can connect directly without a separate process in between. Authentication is a bearer token from the plugin's settings. Because the server lives inside the running app, it can act on live state that a folder read cannot see.

Filesystem servers. These read and write .md files directly and never touch the application. Obsidian does not need to be installed, let alone open, which makes them the option that works on a server or inside a scheduled job. Recent versions added handling for .base and .canvas files as those formats spread through vaults.

Semantic bridges. Retrieval by default is file reads and keyword matching. Plugins that embed the vault on-device — Smart Connections most prominently — can be exposed through their own MCP servers, giving an agent meaning-based retrieval at block level without sending vault contents to a third party.

Headless operation. Obsidian shipped an official headless Sync client in February 2026, which keeps a vault current on a machine with no GUI. Paired with a filesystem server, this is how a vault becomes reachable from a cloud agent or a phone — at the cost of a server the user now administers.

Why It Matters

Obsidian MCP was the first widely adopted way to give an AI agent persistent, user-controlled memory, and it demonstrated demand for the category before any vendor built for it. Its architecture is also the clearest illustration of what a local-first tool can and cannot do for an agent: retrieval quality is solvable by plugin, but whole-vault exposure and the requirement that some machine stay running are properties of the design rather than gaps in it.

The pattern's limits are what purpose-built systems are built against. Permission-aware retrieval, hosted access with no local process, and separation between human-authored and agent-authored content are the differences that show up once a vault is large enough or private enough to matter.

FAQ

Is there an official Obsidian MCP server?

No. Every Obsidian MCP server is community-built. Obsidian's roadmap lists no AI or MCP features, and the company has left the surface to its plugin ecosystem.

Does Obsidian need to be running?

Only for in-app servers. The Local REST API path requires the application open; filesystem servers read the vault folder directly and work with Obsidian closed or not installed.

Can an Obsidian MCP server work from mobile or a web client?

Only through infrastructure the user provides. Obsidian's headless Sync client can hold a vault on a server, and a filesystem MCP server there can be reached remotely — but that server is the user's to run, patch, and secure.

Can an agent be restricted to part of a vault?

Not natively. A server is configured with a vault path and has whatever access that path allows. Scoping requires splitting vaults or running read-only, both of which cost functionality.

How does an agent search a vault semantically?

Through a plugin. Smart Connections generates on-device embeddings and a companion MCP server exposes them, which is meaningfully better than keyword file reads but adds two components to install and maintain.

Related

Contexts

Created with 💜 by One Inc | Copyright 2026