MCP Director

Centralized hub for managing MCP (Model Context Protocol) servers. Acts as a proxy/aggregator, providing a single entry point for AI clients (Claude Desktop, Claude Code, Cursor, etc.).

Architecture

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│  Claude Desktop │     │   Claude Code   │     │     Cursor      │
└────────┬────────┘     └────────┬────────┘     └────────┬────────┘
         │                       │                       │
         │ STDIO                 │ STDIO                 │ HTTP/SSE
         │                       │                       │
         └───────────────────────┴───────────────────────┘
                                 │
                    ┌────────────▼────────────┐
                    │      MCP Director       │
                    │  ┌───────────────────┐  │
                    │  │  Router + ACL     │  │
                    │  │  Profiles/Projects│  │
                    │  │  Analytics        │  │
                    │  └───────────────────┘  │
                    └────────────┬────────────┘
                                 │
         ┌───────────────────────┼───────────────────────┐
         │                       │                       │
         ▼                       ▼                       ▼
┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│  MCP Server 1   │     │  MCP Server 2   │     │  MCP Server N   │
│   (filesystem)  │     │    (github)     │     │     (custom)    │
└─────────────────┘     └─────────────────┘     └─────────────────┘

Key Features

• Aggregation: Combine multiple MCP servers into one endpoint
• Transports: Support for STDIO and HTTP (SSE) transports for clients
• Access Control: Flexible profiles with white/black lists (glob patterns), rate limiting, and time restrictions
• Projects: Ability to create different sets of servers for different projects
• Web Interface: Integrated web interface for managing servers and viewing analytics
• Analytics: Logging of all requests with performance metrics

Quick Start

Requirements

• Rust (latest stable)
• Node.js (required for web interface and some MCP servers)

Installation

cargo install --path apps/mcp-director-cli

Docker

docker-compose up -d

Running

mcp-director serve --config config/default.yaml

Global options: --json (machine-readable output), --dry-run (preview changes), --yes (auto-confirm).

Usage with Clients

Single Client Mode

For a single AI client, use serve directly:

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "director": {
      "command": "mcp-director",
      "args": ["serve"]
    }
  }
}

Multi-Client Mode

For multiple AI clients sharing the same MCP Director instance:

Step 1: Start the daemon:

mcp-director serve --no-stdio

Step 2: Configure clients to use connect:

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "director": {
      "command": "mcp-director",
      "args": ["connect", "--project", "my-project"]
    }
  }
}

Claude Code (.mcp.json):

{
  "mcpServers": {
    "director": {
      "command": "mcp-director",
      "args": ["connect", "--project", "my-project"]
    }
  }
}

HTTP Mode (Cursor, etc.)

1. Go to MCP settings in Cursor
2. Add a new server:
   • URL: http://localhost:3000/sse
   • Transport: SSE
   • Optional: append ?project=your-project&profile=your-profile

Documentation

Full documentation is available at ursul.github.io/mcp-director

• Installation
• Configuration
• HTTP API
• OpenAPI Spec
• CLI Reference
• Architecture
• Development
• Testing
• Troubleshooting

Contributing

We welcome contributions! Please see our CONTRIBUTING.md for details on our code of conduct, development workflow, and how to submit pull requests.

License

MIT