MCP Gateway

<p align="center"> <img src="extras/images/logo.png" width="380" alt="MCP Gateway" /> </p>

<p align="center"> A generic, subclassable Model Context Protocol gateway whose tools are registered and executed asynchronously by connected providers over WebSocket. </p>

<p align="center"> <code>MCP Client</code> ⇄ Streamable HTTP ⇄ <code>Gateway</code> ⇄ WebSocket ⇄ <code>Provider</code> </p>

<p align="center"> <a href="https://github.com/mcp-gtw/mcp-gtw/actions/workflows/ci.yml"><img src="https://github.com/mcp-gtw/mcp-gtw/actions/workflows/ci.yml/badge.svg" alt="CI"></a> <a href="https://pypi.org/project/mcp-gtw/"><img src="https://img.shields.io/pypi/v/mcp-gtw.svg" alt="PyPI"></a> <a href="https://pypi.org/project/mcp-gtw/"><img src="https://img.shields.io/pypi/pyversions/mcp-gtw.svg" alt="Python"></a> <a href="LICENSE.md"><img src="https://img.shields.io/badge/license-MIT-blue.svg" alt="License: MIT"></a> </p>

---

The gateway never knows the capabilities ahead of time. A provider — typically a browser app — connects, publishes its own MCP tools, resources, prompts and more, and executes them when an MCP client calls them. The gateway only registers, publishes and routes, staying completely domain agnostic.

It is a small, installable library. You build a real application by subclassing Gateway and overriding hooks. The handlers themselves are written in JavaScript. This library never prescribes a specific frontend — the Provider SDK guide teaches the JS side generically.

📦 Install

pip install mcp-gtw
# or
uv add mcp-gtw

🚀 The smallest gateway

from mcp_gtw.gateway import Gateway

app = Gateway().create_app()

uv run python -m mcp_gtw.main

This publishes a real MCP endpoint at /mcp, a private provider WebSocket at /provider and a health check at /health. It exposes whatever capabilities the connected provider registers.

☁️ One-click deploy

The image runs production-ready as a non-root process. Deploy it to any host that keeps a persistent server (the /provider WebSocket needs one — serverless like Vercel does not work):

Render, Railway and Fly.io run the Dockerfile as-is (they inject PORT, which the gateway reads); on a VPS (Hostinger, …) use Docker Compose. Full guide, including reverse proxy and TLS: deployment.

🧩 Extending it

Subclass Gateway and override the hooks to attach your own domain logic:

from mcp_gtw.channel import Channel
from mcp_gtw.gateway import Gateway

class MyGateway(Gateway):
    mcp_server_name = "my-app"

    async def on_provider_connected(self, channel: Channel) -> None:
        ...  # a provider session just came online

    def register_routes(self, app) -> None:
        super().register_routes(app)
        ...  # add your own HTTP and WebSocket routes

app = MyGateway().create_app()

Every behaviour is a swappable strategy with a secure default — authentication, tokens, origins, expiry and the wire codec. Change one by setting a *_class attribute or injecting an instance, without touching the transport. See the Gateway library guide for every override point, Extensibility for the strategy contracts, and Auth recipes for token, username/password and client-supplied-token models.

📚 Documentation

Guide	What it covers
Architecture	Components, transports and request flows.
Quick start	Install, run and connect an MCP client.
Gateway library	The Gateway class and every override point.
Extensibility	The swappable strategies, their contracts and the invariants.
Auth recipes	Token, username/password and client-supplied-token models.
Configuration	Every setting and environment variable.
Provider protocol	The private gateway ⇄ provider message protocol.
Provider SDK	Writing the JavaScript provider and registering tools.
Browser console	Turn any open page into a provider from DevTools.
MCP clients	Connecting Claude Code, generic clients and the Inspector.
Admin dashboard	The optional monitoring dashboard and its stats API.
Security	The security model, tokens, origins and hardening.
Testing	Running the suite and the 100% coverage gate.
Deployment	Docker, reverse proxies and scaling.

🗂️ Layout

.
├── src/mcp_gtw/     # the library
├── tests/               # unit and integration tests (100% coverage)
└── docs/                # the guides linked above

✅ Requirements

• Python 3.12+ — tested on 3.12, 3.13 and 3.14 in CI (3.12 is the pinned local and Docker version)
• Any MCP client (Claude Code, Cursor, the MCP Inspector, …)

💜 Support

If this project saved you time, consider supporting it: GitHub Sponsors · Ko-fi.

Made with care by Paulo Coutinho.

Licensed under MIT.