Manhattan MCP (GitMem)

Token-Efficient Codebase Navigation - MCP Server for the Manhattan Memory System.

Manhattan MCP is a local Model Context Protocol (MCP) server that provides AI agents (Claude Desktop, Cursor, Windsurf, etc.) with a Virtual File System (VFS) backed by compressed, cached code context. It allows agents to understand large codebases while saving 50-80% on tokens.

Features

• 🏗️ GitMem Context - Compressed semantic code skeletons (signatures, summaries, hierarchies).
• 🔍 Hybrid Search - Semantic and keyword search across the entire codebase.
• 📂 VFS Navigation - Browse and read files via token-efficient outlines and contexts.
• 📝 Auto-Indexing - Automatically keeps the code index fresh after edits.
• 📊 Token Analytics - Track token savings and repository indexing status.

Installation

pip install manhattan-mcp

Quick Start

1. Install Manhattan MCP

pip install manhattan-mcp

2. Configure Your AI Client

Setting up Manhattan MCP is a two-step process:

Step A: Register the Server (One-time)

Add Manhattan MCP to your AI tool's global settings.

Claude Desktop Add to claude_desktop_config.json:

{
  "mcpServers": {
    "manhattan": {
      "command": "manhattan-mcp",
      "args": ["start"]
    }
  }
}

Cursor Add to Cursor Settings > MCP:

• Name: manhattan
• Type: command
• Command: manhattan-mcp

GitHub Copilot (VS Code) Add to your Copilot MCP settings:

{
  "servers": {
    "manhattan": {
      "command": "manhattan-mcp",
      "args": ["start"]
    }
  }
}

Step B: Apply Project Rules (Per Project)

Run the setup command in your project root to ensure the agent follows the mandatory indexing policy.

# For Cursor
manhattan-mcp setup cursor

# For Claude
manhattan-mcp setup claude

# For Gemini (Antigravity)
manhattan-mcp setup gemini

# For GitHub Copilot
manhattan-mcp setup copilot

# For Windsurf
manhattan-mcp setup windsurf

# For all supported clients
manhattan-mcp setup all

3. Start Navigating!

Once configured, your AI agent can use Manhattan MCP to understand your codebase efficiently.

Example Usage

Searching for code:

User: How does the authentication flow work?
AI: *calls search_codebase "authentication flow"*
    I found the authentication logic in `auth.py`. 
    Let me read the context for you...

Understanding a file:

User: Summarize the main functions in server.py
AI: *calls get_file_outline "src/server.py"*
    The server.py file contains:
    - `start_server()`: Initializes the FastMCP instance...
    - `api_usage()`: Returns usage statistics...

Saving tokens:

User: Read the implementation of the memory builder.
AI: *calls read_file_context "src/builder.py"*
    (Returns a compressed semantic skeleton, saving 70% tokens)
    The memory builder uses a two-phase ingestion process...

Configuration Options

CLI Commands

# Start the MCP server (default)
manhattan-mcp start

# Set up client rules (Cursor, Claude, etc.)
manhattan-mcp setup [client]

# Show version
manhattan-mcp --version

# Show help
manhattan-mcp --help

License

MIT License - see LICENSE for details.

Links

• 🌐 Website
• 📖 Documentation
• 🐛 Issues
• 💬 Discord