MetaMCP Local Proxy
metatool-aiLABSCO SUMMARY
One connection instead of twelve, with containers and stored secrets in between.
The part that saves real time is adding a server by URL and letting the analysis work out how to run it, rather than reading someone's README to find the launch command. Secret files are handled properly — encrypted at rest, mounted read-only, and never returned in profile listings — which is the thing that usually stops teams from sharing a server setup.
What it is
A gateway that installs MCP servers from GitHub repositories, keeps their credentials, and re-exposes the tools you enable through a single connection.
What you get
- Add a server by pasting a GitHub URL — the repository is analyzed and configured automatically
- Each server runs in its own Docker container, with npx, uvx and Docker runtimes supported
- Per-server tool enablement, so a client sees only the tools you switched on
- Centralized API keys, plus encrypted secret files — JSON credentials, PEM certificates, PKCS12, Java KeyStores — mounted read-only into the container and exposed as a path in an environment variable you name
- Start, stop and restart servers from the web interface, with container health monitoring and remounting
- Google OAuth sign-in with optional domain restriction, session auth for the web UI, and JWT tokens for API access
- Runs as stdio for a local desktop client, or over HTTP with OAuth for a shared deployment
Requirements
SECRET_KEY and ANTHROPIC_API_KEY — the Anthropic key is what analyzes repositories on add. Docker is needed for the isolation. Install with uv sync or pip, then start with mcp-anywhere serve http and open port 8000, or mcp-anywhere serve stdio for a local desktop client. Google sign-in is optional and needs GOOGLE_OAUTH_CLIENT_ID and GOOGLE_OAUTH_CLIENT_SECRET. The project describes itself as beta.
Setup effort
Build from source — git clone https://github.com/locomotive-agency/mcp-anywhere.git
USE CASES
01
Cutting a long list of servers down to one connection
Register each server once in the gateway; your client connects to a single endpoint and sees the tools you enabled.
02
Trying out a server you just found on GitHub
Paste the repository URL and let the analysis work out the runtime and configuration instead of reading the README.
03
Sharing a tool setup with a team
Run it over HTTP with Google sign-in restricted to your domain, so colleagues authenticate rather than copying config files.
04
Handing a server a credential file safely
Upload the JSON or PEM, name the environment variable, and it is mounted read-only into that server's container only.
05
Keeping one noisy server from touching the rest
Each server runs in its own container, and you can stop or restart just that one from the web interface.
BEFORE YOU INSTALL
What to know first
Requires
ANTHROPIC_API_KEY in your environment
Where it runs
The documented client runs the server on your machine.
INSTALLATION
Connect your client
Server setup
TERMINAL
# Clone repository
git clone https://github.com/locomotive-agency/mcp-anywhere.git
cd mcp-anywhere
# Install with uv (recommended)
uv sync
# Or install with pip
pip install -e .
# Configure environment
cp env.example .env
# Edit .env with required values:
# SECRET_KEY=<YOUR_SECRET_KEY>
# ANTHROPIC_API_KEY=<YOUR_ANTHROPIC_API_KEY>
# Start server
mcp-anywhere serve http
# Or: python -m mcp_anywhere serve http
# Access at http://localhost:8000MCP Clients
Claude Desktop
CLAUDE DESKTOP · JSON
{
"mcpServers": {
"mcp-anywhere": {
"command": "/Users/<YOUR_USERNAME>/.local/bin/uv",
"args": [
"run",
"--directory",
"/Users/<YOUR_USERNAME>/Projects/mcp-anywhere",
"mcp-anywhere",
"connect"
]
}
}
}The uv path is the one which uv prints, and the --directory path is the path to your MCP Anywhere clone. Set up the application with uv run mcp-anywhere serve stdio before starting Claude Desktop.
