About mcp-gateway
mcp-gateway is a hub that connects downstream MCP servers and presents their tools through a unified search and calling interface. It is suited to setups that need deferred tool loading, configurable authentication, and stdio or HTTP transport.
Key features
- 01Register downstream servers through JSON config
- 02Namespace tools by server
- 03Search with BM25 or regex
- 04Defer tool loading until needed
- 05Support stdio and HTTP transports
- 06Cache tool definitions
Use cases
- 01Search tools across several MCP servers
- 02Load only relevant tools to limit context use
- 03Call remote tools with configured authentication
MCP support
It connects to downstream MCP servers, discovers and indexes their tools, and exposes search and remote-call tools to MCP clients.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
uvx mcp-orchestration-gatewaySTORAGE_BACKENDStorage backend for tool cache. Use 'memory' for development or 'redis' for production.REDIS_URLRedis connection URL. Required only when STORAGE_BACKEND=redis.SERVER_CONFIG_PATHPath to the server_config.json file listing downstream MCP servers.Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.
Tools
Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.
Details
From the project READMEMCP Orchestrator
A central hub that connects to multiple downstream MCP servers, aggregates their tools, and provides unified access with powerful tool search capabilities.
Built around deferred tool loading — search across all your servers without blowing Claude's context window.
Features
- Config-based Server Registration: Add downstream MCP servers via JSON config file
- Tool Namespacing: Automatic
server_name__tool_nameformat - Tool Search: Unified BM25/regex search with deferred loading support
- Flexible Authentication: Static saved headers or token forwarding
- Multiple Transports: stdio or HTTP
- Tool Definition Caching: Cached definitions, raw result passthrough
- Storage Backends: In-memory (development) or Redis (production)
Quick Start
Installation
pip install mcp-orchestrator
Running the MCP Server
# Run as stdio MCP server (for Claude Desktop, Cursor, etc.)
mcp-orchestrator
# Or run with Python directly
python -m mcp_orchestrator.main
HTTP Transport:
ORCHESTRATOR_TRANSPORT=http ORCHESTRATOR_PORT=8080 python -m mcp_orchestrator.main
This starts the server on http://localhost:8080/mcp with CORS enabled.
Configuring Servers
Add downstream MCP servers in server_config.json:
{
"servers": [
{
"name": "my-server",
"url": "http://localhost:8080/mcp",
"transport": "http",
"auth_type": "static",
"auth_headers": {
"Authorization": "Bearer my-token"
}
},
{
"name": "my-stdio-server",
"url": "server.py",
"transport": "stdio",
"command": "uv",
"args": ["run", "python", "server.py"]
}
]
}
Searching for Tools
The orchestrator provides unified tool search (BM25 by default, regex optional):
# BM25 search (default - natural language)
results = await mcp_client.call_tool("tool_search", {
"query": "get weather information",
"max_results": 3
})
# Regex search (set use_regex=true)
results = await mcp_client.call_tool("tool_search", {
"query": "weather|forecast",
"use_regex": true,
"max_results": 3
})
Architecture
┌─────────────────────────────────────────────────────┐
│ MCP Orchestrator │
│ │
│ ┌──────────────────────────────────────────────┐ │
│ │ FastMCP Server │ │
│ │ ┌─────────────┐ ┌──────────────────┐ │ │
│ │ │ tool_search │ │ call_remote_tool │ │ │
│ │ └─────────────┘ └──────────────────┘ │ │
│ └──────────────────────────────────────────────┘ │
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────────┐ │
│ │ Server │ │ Tool │ │ Storage │ │
│ │ Registry │ │ Search │ │(Memory/Redis)│ │
│ └──────────┘ └──────────┘ └──────────────┘ │
└─────────────────────────────────────────────────────┘
│
┌───────────────────┼───────────────────┐
▼ ▼ ▼
┌─────────┐ ┌─────────┐ ┌─────────┐
│ MCP Svr │ │ MCP Svr │ │ MCP Svr │
│ #1 │ │ #2 │ │ #N │
└─────────┘ └─────────┘ └─────────┘
Configuration
Environment Variables
| Variable | Default | Description |
|---|---|---|
STORAGE_BACKEND | memory | Storage backend (memory or redis) |
REDIS_URL | redis://localhost:6379/0 | Redis connection URL |
MCP_ORCHESTRATOR_TOOL_CACHE_TTL | 300 | Tool schema cache TTL in seconds |
MCP_ORCHESTRATOR_DEFAULT_CONNECTION_MODE | stateless | Default connection mode |
MCP_ORCHESTRATOR_CONNECTION_TIMEOUT | 30.0 | Connection timeout in seconds |
MCP_ORCHESTRATOR_MAX_RETRIES | 3 | Maximum retry attempts |
ORCHESTRATOR_TRANSPORT | stdio | MCP transport (stdio or http) |
ORCHESTRATOR_PORT | 8080 | Port for HTTP transport |
ORCHESTRATOR_HOST | 0.0.0.0 | Host for HTTP transport |
ORCHESTRATOR_LOG_LEVEL | INFO | Logging level |
SERVER_CONFIG_PATH | server_config.json | Path to server configuration file |
Claude Desktop Integration
Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):
{
"mcpServers": {
"mcp-orchestrator": {
"command": "mcp-orchestrator",
"env": {
"STORAGE_BACKEND": "memory",
"ORCHESTRATOR_LOG_LEVEL": "INFO"
}
}
}
}
MCP Tools
tool_search
Search for tools using BM25 relevance ranking or regex pattern matching.
@mcp.tool()
async def tool_search(
query: str,
max_results: int = 3,
use_regex: bool = False,
) -> dict:
"""Search for tools using BM25 or regex.
By default uses BM25 natural language search. Set use_regex=True
to search using Python regex patterns instead.
"""
discover_tools
Discover tools from a registered downstream server.
@mcp.tool()
async def discover_tools(
server_name: str,
) -> dict:
"""Discover tools from a registered server and index them for search.
Returns the list of discovered tools with their schemas.
"""
call_remote_tool
Call a tool directly on a downstream MCP server.
@mcp.tool()
async def call_remote_tool(
tool_name: str,
arguments: Optional[dict] = None,
auth_header: Optional[str] = None,
) -> Any:
"""Call a tool on a downstream server.
Args:
tool_name: Namespaced tool name (server_name__tool_name)
arguments: Tool arguments
auth_header: Optional auth header to override server's configured auth
"""
Tool Search Results
The search tools return results in the format expected by Claude's tool search system:
{
"success": true,
"tool_references": [
{
"type": "tool_reference",
"tool_name": "server_name__tool_name"
}
],
"total_matches": 5,
"query": "weather"
}
Testing
Run the test suite:
uv run pytest
Run with coverage:
uv run pytest --cov=mcp_orchestrator
Project Structure
mcp-orchestrator/
├── src/mcp_orchestrator/
│ ├── __init__.py
│ ├── main.py # Entry point
│ ├── models.py # Pydantic models
│ ├── mcp_server.py # FastMCP server
│ ├── config_loader.py # Config file loader
│ ├── server/
│ │ └── registry.py # Server registry
│ ├── tools/
│ │ ├── router.py # Tool router
│ │ └── search.py # Tool search service
│ └── storage/
│ ├── base.py # Storage interface
│ ├── memory.py # In-memory backend
│ └── redis.py # Redis backend
├── tests/
│ ├── test_registry.py
│ ├── test_search.py
│ ├── test_storage.py
│ ├── test_models.py
│ └── test_integration.py
├── server_config.json # Pre-configured downstream servers
├── pyproject.toml
├── README.md
└── .env # Environment variables (not committed)
License
MIT License
Contributing
Contributions are welcome! Please see CONTRIBUTING.md for guidelines.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.