MCP 服务器

mcp-recall-md

Enables semantic search over local markdown note collections using vector embeddings, with real-time file watching and zero-config setup.

README

<h1 align="center">mcp-recall-md</h1>

Local semantic search for your markdown notes — via <a href="https://modelcontextprotocol.io/">MCP</a>.

graph LR
    A["<b>.md files</b><br><i>Obsidian · Logseq · plain md</i>"] -- "watch + index" --> B["<b>mcp-recall-md</b><br><i>local vector store</i>"]
    C["<b>AI client</b><br><i>Claude · Cursor · Windsurf</i>"] -- "MCP search" --> B
    style A fill:#f9f4eb,stroke:#c9a96e,color:#333
    style B fill:#e8f0fe,stroke:#4a86c8,color:#333
    style C fill:#e8f5e9,stroke:#5b9a5b,color:#333

"Search my notes about Kubernetes networking"

→ finds kubernetes-networking.md (similarity: 0.53) — even though you phrased it differently than the note

Search by meaning, not keywords — finds notes even when your wording doesn't match
100% offline — no API keys, no cloud, nothing leaves your machine
Zero config — point at your folders, restart your AI client, done
Real-time sync — file watcher picks up changes instantly, re-embeds only what changed

Quick start

Add to your MCP client config and restart:

{
  "mcpServers": {
    "mcp-recall-md": {
      "command": "uvx",
      "args": ["mcp-recall-md", "--vaults", "C:/Users/you/notes"]
    }
  }
}

Config file location: .mcp.json (Claude Code) · claude_desktop_config.json (Claude Desktop) · Cursor / Windsurf MCP settings

That's it. Your notes are searchable.

Installation

The quick start above uses uvx (recommended). Other options:

pip install mcp-recall-md

{
  "mcpServers": {
    "mcp-recall-md": {
      "command": "mcp-recall-md",
      "args": ["--vaults", "C:/Users/you/notes"]
    }
  }
}

</details>

<details> <summary>Standalone .exe (no Python needed)</summary>

Download mcp-recall-md.exe from the latest release
Put it somewhere permanent (e.g. C:\Tools\mcp-recall-md\)

{
  "mcpServers": {
    "mcp-recall-md": {
      "command": "C:/Tools/mcp-recall-md/mcp-recall-md.exe",
      "args": ["--vaults", "C:/Users/you/notes"]
    }
  }
}

</details>

Configuration

Multiple vaults

List all folders — each is indexed independently:

"args": ["--vaults", "C:/notes/work", "C:/notes/personal", "C:/docs"]

.recallignore

Drop a .recallignore in any vault root to exclude files. Standard .gitignore syntax:

.obsidian/
_templates/
drafts/

CLI flags

Flag	Default	Description
`--vaults`	(none)	Folders to index and watch
`--db-path`	`~/.mcp-recall-md/db`	ChromaDB storage location
`--verbose`	off	Debug logging to stderr

Tools

Your AI assistant gets these tools automatically via MCP:

Tool	Description
`search`	Find notes by meaning — returns ranked results with similarity scores and file paths
`status`	Show indexed article count and watched vaults
`index`	Manually store an article (for use without `--vaults`)
`remove`	Delete an article from the index

Most users only interact with search — everything else is automatic.

Troubleshooting

Problem	Fix
Search returns nothing	Check that `--vaults` points to folders with `.md` files
First run is slow	Embedding model (~80 MB) downloads once on first use
Need to debug	Add `--verbose`, check `~/.mcp-recall-md/server.log`
Force re-index	Delete `~/.mcp-recall-md/db` and restart

Logs: ~/.mcp-recall-md/server.log (5 MB max, 3 rotated backups)

Limitations

Single-chunk embedding — large files (10k+ words) may search less precisely than shorter notes
English-optimized — other languages work but with lower accuracy