Obsidian MCP: connect Claude to your vault (2026)

The quickest way to connect Claude to an Obsidian vault in 2026 is the Local REST API community plugin, which has had a built-in MCP server since version 4.0.0. Install it, copy its API key, and add the server to Claude Code with one command. If you would rather not run a plugin, MCPVault reads the vault folder directly and works while Obsidian is closed. We make Loam, a Mac notebook that has no MCP server, and we explain the file-and-command-line route at the end.

What is an Obsidian MCP server?

MCP, the Model Context Protocol, is an open standard that lets an AI assistant call outside tools. An Obsidian MCP server exposes your vault as tools: list files, read a note, search, append text, patch one heading, move or delete a file. The assistant picks a tool, the server runs it against your notes, and the result goes back into the conversation.

Obsidian does not provide this itself. We looked through the official help pages on 2026-11-05 and found a command line interface, headless sync and publish tools, and no MCP page. Every MCP route below comes from a community project, so treat each one as third-party code.

Which Obsidian MCP options exist in 2026?

Five routes come up most often. The first three are real MCP servers. The fourth is the official command line, which is not MCP but solves the same problem for a coding agent. The fifth is a plugin you will still see recommended, and its author has stopped maintaining it.

How do you connect Claude Code to Obsidian?

The built-in server of the Local REST API plugin is the shortest path, because Claude Code speaks HTTP MCP natively. Follow these steps.

• In Obsidian, open Settings, then Community plugins, and install and enable Local REST API.
• Open Settings, then Local REST API, and copy the API key.
• Run the command below in a terminal, with your key in place of the placeholder.
• Start Claude Code and ask it to list the files in your vault.

claude mcp add --transport http obsidian https://127.0.0.1:27124/mcp/ \
  --header "Authorization: Bearer <your-api-key>"

Two details trip people up. First, the plugin signs its endpoint with a certificate authority it generates on your machine, so your client must trust that certificate. You can download it from https://127.0.0.1:27124/obsidian-local-rest-api.crt. If trusting it is not possible, switch on Enable HTTP server in the plugin settings and use http://127.0.0.1:27123/mcp/ instead. Second, Claude Code stores a server in its local scope by default. Add --scope user to use the vault from every project, or --scope project to share it through a .mcp.json file. Because that file would contain your key, keep the key out of version control.

Run claude mcp list afterwards. It shows a connected or failed status next to each server.

How do you connect Claude Desktop to Obsidian?

Claude Desktop does not connect to remote HTTP servers on its own, according to the plugin's README, so you have two options. You can bridge the plugin with mcp-remote, or you can skip the plugin and let MCPVault read the vault folder. The second one is simpler, because it is a single entry in claude_desktop_config.json. Replace the path with your vault, save the file and restart the app.

{
  "mcpServers": {
    "obsidian": {
      "command": "npx",
      "args": ["@bitbonsai/mcpvault@latest", "/path/to/your/vault"]
    }
  }
}

MCPVault needs Node.js 22.12.0 or later. The project lists configuration examples for Claude Code, ChatGPT Desktop (Enterprise and above), OpenCode, Gemini CLI, Codex, Cursor, Windsurf and IntelliJ IDEA, so the same server works across clients.

How do you set up Obsidian MCP for Cursor?

Cursor supports the streamable HTTP transport, so it can use the Local REST API server directly. Add this to ~/.cursor/mcp.json for every project, or to .cursor/mcp.json for one project.

{
  "mcpServers": {
    "obsidian": {
      "url": "https://127.0.0.1:27124/mcp/",
      "headers": {
        "Authorization": "Bearer <your-api-key>"
      }
    }
  }
}

What can the agent do once it is connected?

The Local REST API server groups its 20 tools into a few families. Reading covers vault_list, vault_read, vault_get_document_map (headings, block references and frontmatter fields of a file) and vault_read_binary for images and attachments. Writing covers vault_write, vault_append, vault_patch for a single heading, block or frontmatter field, plus vault_move, vault_copy and vault_delete. Search comes in two forms: search_simple, which uses Obsidian's own search, and search_query, which filters note metadata with JsonLogic. The remaining tools list tags, open a file in the app, and list or run Obsidian commands.

command_execute lets the agent run any command from your command palette, including commands added by other plugins, so its reach goes beyond note editing. MCPVault is narrower: its 18 tools work on files, frontmatter, tags and wiki links, and it never touches the app. It also offers search_notes with BM25 ranking, and read_note_lines and get_note_outline so the agent can read part of a long note instead of all of it.

Search results go straight into the model's context. The plugin's README warns that a plain search with common words could return megabytes, so search_simple now returns at most 20 files and 10 matches per file unless the call asks for more. Ask for targeted questions, and let the agent read an outline before it opens a long note.

Is the Obsidian MCP server safe?

The risk depends on the access you grant, and the access is broad. Anything that holds your API key can read, change and delete notes. Release 5.4.0 of the plugin fixed a vulnerability that let an authenticated client write into the .obsidian folder and alter plugin code, and the API now refuses that folder unless you turn on an advanced setting. Update to 5.4.0 or later. These habits keep the rest under control.

• Put the vault under Git and commit before each agent session. A diff shows exactly what the agent changed, and one command undoes it.
• Prefer a copy of the vault for the first experiments, then switch to the real one.
• Keep the API key out of files you commit, such as a project-level .mcp.json.
• Read the source of any third-party server before you run it. MCPVault is MIT licensed and restricts file operations to the vault root, and it blocks .obsidian, .git and node_modules. Its delete_note and move_file tools require a matching confirmation path.
• Expect deletes in the REST API to go to Obsidian's trash by default. Since 5.5.0 every write accepts If-Match, so two clients editing one note cannot silently overwrite each other.

Should you use Obsidian MCP or the Obsidian CLI?

Use the CLI when your agent has a terminal, and MCP when it does not. Obsidian's command line covers search, reading, creating notes, daily notes, tasks, tags and even developer commands such as running JavaScript in the app console. You switch it on under Settings, then General, then Command line interface. It needs the 1.12 installer (1.12.7 or later) and the app has to be running, although the first command launches it.

obsidian daily:append content="- [ ] Review the roadmap"
obsidian search query="meeting notes"
obsidian create name="Trip to Paris" template=Travel

A coding agent such as Claude Code or Codex can run these commands with no server to configure. A chat assistant without a shell, such as Claude Desktop in a plain chat, has no way to call them, and that is where an MCP server earns its place.

How does Loam compare with an Obsidian MCP server?

Loam is a free, open source Markdown notebook for macOS. It keeps your notes as plain .md files in a folder, as Obsidian does, and it takes the other route: no server at all. An agent works on the vault through the loam command line, and it reads the rules for doing so from loam guide. Install the skill with loam skill install --claude, and Claude Code follows those rules before it touches a note.

Where does Loam fall short?

Loam has no MCP server. An assistant that can only speak MCP and has no terminal cannot reach a Loam vault directly. Loam also runs only on a Mac with Apple silicon and macOS 26 or later, while Obsidian covers more platforms and has a far larger plugin catalog and community. If you need Windows, Linux or a phone, or a chat assistant with no shell, an Obsidian MCP route is the better answer. Our Loam vs Obsidian page lists the other differences.

Which route should you choose?

Choose the Local REST API MCP server if:

• You want the agent to see live Obsidian state, such as the active file, tags and commands.
• You are fine with Obsidian staying open while the agent works.
• You can trust a local certificate or enable the plain HTTP endpoint.

Choose MCPVault if:

• You want a server that works with Obsidian closed, and no plugin to maintain.
• You use Claude Desktop, Cursor or Codex and want one setup that works everywhere.
• You prefer a narrow tool set that touches files only.

Choose the Obsidian CLI if your agent already has a terminal and you want the official, supported route. Choose Loam if you work on a Mac, want open source code, and are happy to have your agents use commands rather than MCP. For a wider look at the field, read Note-taking apps with MCP servers and How to use Claude Code with your notes.

Sources: Local REST API for Obsidian: README and MCP section ↗ · Local REST API for Obsidian: releases ↗ · MCPVault: README ↗ · mcp-obsidian: README ↗ · MCP Tools for Obsidian: README (archived) ↗ · Obsidian: Command line interface ↗ · Claude Code: Connect Claude Code to tools via MCP ↗

Try Loam

Loam is free and open source. Install it with one command, run loam skill install --claude, and let your agent work on a copy of your vault. The Agents page lists the rules every agent follows, and the command line reference lists every command. You can also read how Loam stacks up against Obsidian, or see every comparison.