mcp-recall-md 0.2.0

Local semantic search for markdown notes — MCP server

mcp-recall-md

Local semantic search for your markdown notes — via MCP
Search by meaning, not keywords. 100% offline.

Install  ·   Tools  ·   Auto-indexing  ·   .recallignore  ·   Troubleshooting

---

Install

Option A: Python users

pip install mcp-recall-md

Add to your MCP config (.mcp.json for Claude Code, claude_desktop_config.json for Claude Desktop):

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

Option B: Download exe (no Python needed)

1. Download mcp-recall-md.exe from the latest release
2. Put it in a permanent folder (e.g. C:\Tools\mcp-recall-md\)
3. Add to MCP config:

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

Multiple vaults? Just list them:

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

Restart your app. Start asking:

"Search my knowledge base for Kubernetes networking"

"Find my notes about AWS Bedrock"

---

Tools

Tool	Parameters	What it does
index	key, content	Store or update an article
search	query, limit	Semantic similarity search
remove	key	Delete an article

---

Auto-indexing

When --vaults is provided, the server automatically:

1. Indexes all existing .md files on startup
2. Watches for new and modified files in real time

No separate watcher process needed — it's all one command.

Without --vaults, the server runs in manual mode (use the index tool to add content).

---

.recallignore

Drop a .recallignore file in any vault root to control which files get indexed. Uses standard .gitignore syntax.

Exclude specific folders:

.obsidian/
_templates/
drafts/

Include only specific folders (exclude everything else):

*
!notes/
!docs/

Each vault gets its own .recallignore — they're independent.

---

Troubleshooting

Problem	Fix
Search returns nothing	Make sure --vaults is set, or use the index tool manually
First run is slow	Embedding model (~80 MB) downloads once
Need to debug	Add --verbose flag

---

License

MIT