mcp-engine

MCPEngine

Production-Grade Implementation of the Model Context Protocol (MCP)

Overview

MCPEngine is a production-grade, HTTP-first implementation of the Model Context Protocol (MCP). It provides a secure, scalable, and modern framework for exposing data, tools, and prompts to Large Language Models (LLMs) via MCP.

We believe MCP can be the "REST for LLMs," enabling any application (Slack, Gmail, GitHub, etc.) to expose a standardized endpoint that LLMs can access without custom-coded integrations. MCPEngine is our contribution to making MCP robust enough for modern, cloud-native use cases.

Key Features

• Built-in OAuth with Okta, Keycloak, Google SSO, etc.
• HTTP-first design (SSE instead of just stdio)
• Scope-based Authorization for tools, resources, and prompts
• Seamless bridging for LLM hosts (like Claude Desktop) via a local proxy
• Full backwards-compatibility with FastMCP and the official MCP SDK

Architecture

MCPEngine uses a proxy-based architecture to integrate with LLM hosts like Claude Desktop:

┌───────────────┐     stdio     ┌─────────────────┐     HTTP/SSE     ┌───────────────┐
│  Claude Host  ├───────────────►  MCPProxy Local ├──────────────────► MCPEngine     │
│               │               │                 │                   │ Server        │
│               ◄───────────────┤ (runs locally) ◄──────────────────┬┤ (remote)      │
└───────────────┘               └─────────────────┘      OAuth 2.1   │└───────────────┘
                                                                     │
                                                        ┌────────────┴───────────┐
                                                        │ Identity Provider      │
                                                        │ (Okta, Keycloak, etc.) │
                                                        └────────────────────────┘

This architecture provides several advantages:

1. Seamless integration - Claude sees a local stdio-based process
2. Security - The proxy handles OAuth authentication flows
3. Scalability - The MCPEngine server can run anywhere (cloud, on-prem)
4. Separation of concerns - Authentication is handled independently from your business logic

Installation

uv add "mcpengine[cli]"
# or
pip install "mcpengine[cli]"

Once installed, you can run the CLI tools:

mcpengine --help

Requirements: in order to run the proxy provided by MCPEngine (see below), the Docker daemon needs to be & up and running on local system with access to Docker Hub to pull image featureformcom/mcpengine-proxy:latest.

Quickstart

Create a Server

# server.py
from mcpengine import MCPEngine

mcp = MCPEngine("Demo")


@mcp.tool()
def add(a: int, b: int) -> int:
    return a + b


@mcp.resource("greeting://{name}")
def get_greeting(name: str) -> str:
    return f"Hello, {name}!"

Claude Desktop Integration

If your server is at http://localhost:8000, you can start the proxy locally:

mcpengine proxy http://localhost:8000/sse

Claude Desktop sees a local stdio server, while the proxy handles any necessary OAuth or SSE traffic automatically.

Core Concepts

Authentication & Authorization

Enable OAuth and scopes:

from mcpengine import MCPEngine, Context
from mcpengine.server.auth.providers.config import IdpConfig

mcp = MCPEngine(
    "SecureDemo",
    idp_config=IdpConfig(
        issuer_url="https://your-idp.example.com/realms/some-realm",
    ),
)


@mcp.auth(scopes=["calc:read"])
@mcp.tool()
def add(a: int, b: int, ctx: Context) -> int:
    ctx.info(f"User {ctx.user_id} with roles {ctx.roles} called add.")
    return a + b

Any attempt to call add requires the user to have calc:read scope. Without it, the server returns 401 Unauthorized, prompting a login flow if used via the proxy.

Resources

@mcp.resource("uri"): Provide read-only context for LLMs, like a GET endpoint.

from mcpengine import MCPEngine

mcp = MCPEngine("Demo")


@mcp.resource("config://app")
def get_config() -> str:
    return "Configuration Data"

Tools

@mcp.tool(): LLM-invokable functions. They can have side effects or perform computations.

from mcpengine import MCPEngine

mcp = MCPEngine("Demo")


@mcp.tool()
def send_email(to: str, body: str):
    return "Email Sent!"

Prompts

@mcp.prompt(): Reusable conversation templates.

from mcpengine import MCPEngine

mcp = MCPEngine("Demo")


@mcp.prompt()
def debug_prompt(error_msg: str):
    return f"Debug: {error_msg}"

Images

Return images as first-class data:

from mcpengine import MCPEngine, Image

mcp = MCPEngine("Demo")


@mcp.tool()
def thumbnail(path: str) -> Image:
    # ... function body omitted
    pass

Context

Each request has a Context:

• ctx.user_id: Authenticated user id
• ctx.user_name: Authenticated user name
• ctx.roles: User scopes/roles
• ctx.info(...): Logging
• ctx.read_resource(...): Access other resources

Example Implementations

SQLite Explorer

import sqlite3
from mcpengine import MCPEngine, Context
from mcpengine.server.auth.providers.config import IdpConfig

mcp = MCPEngine(
    "SQLiteExplorer",
    idp_config=IdpConfig(
        issuer_url="https://your-idp.example.com/realms/some-realm",
    ),
)


@mcp.auth(scopes=["database:read"])
@mcp.tool()
def query_db(sql: str, ctx: Context) -> str:
    conn = sqlite3.connect("data.db")
    try:
        rows = conn.execute(sql).fetchall()
        ctx.info(f"User {ctx.user.id} executed query: {sql}")
        return str(rows)
    except Exception as e:
        return f"Error: {str(e)}"

Echo Server

from mcpengine import MCPEngine

mcp = MCPEngine("Demo")


@mcp.resource("echo://{msg}")
def echo_resource(msg: str):
    return f"Resource echo: {msg}"


@mcp.tool()
def echo_tool(msg: str):
    return f"Tool echo: {msg}"

Smack - Message Storage Example

Smack is a simple messaging service example with PostgreSQL storage that demonstrates MCPEngine's capabilities with OAuth 2.1 authentication.

Quick Start

1. Start the service using Docker Compose:

git clone https://github.com/featureform/mcp-engine.git
cd mcp-engine/examples/servers/smack
docker-compose up --build

2. Using Claude Desktop

Configure Claude Desktop to use Smack:

Manually:

touch ~/Library/Application\ Support/Claude/claude_desktop_config.json

Add to the file:

{
  "mcpServers": {
    "smack_mcp_server": {
      "command": "bash",
      "args": [
        "docker attach mcpengine_proxy || docker run --rm -i --net=host --name mcpengine_proxy featureformcom/mcpengine-proxy -host=http://localhost:8000 -debug -client_id=optional -client_secret=optional",
      ]
    }
  }
}

Via CLI:

mcpengine proxy http://localhost:8000

Smack provides two main tools:

• list_messages(): Retrieves all messages
• post_message(message: str): Posts a new message

For more details, see the Smack example code.

Roadmap

• Advanced Auth Flows
• Service Discovery
• Fine-Grained Authorization
• Observability & Telemetry
• Ongoing FastMCP Compatibility

Contributing

We welcome feedback, issues, and pull requests. If you'd like to shape MCP's future, open an issue or propose changes on GitHub. We actively maintain MCPEngine to align with real-world enterprise needs.

Community

Join our discussion on Slack to share feedback, propose features, or collaborate.

License

Licensed under the MIT License. See LICENSE for details.