MCP Agent

🚀 MCP Proxy

MCP-Proxy is a framework designed to simplify the construction of AI agents by leveraging the capabilities provided by MCP (Model Context Protocol) servers. MCP is a low-level protocol, and this framework takes care of tasks such as connecting to the server, interacting with large language models (LLMs), handling external signals (e.g., human input), and supporting persistent states, allowing you to focus on business logic.

✨ Features

• Interoperability: Ensures that any tools provided by MCP servers can be seamlessly integrated into your agents.
• Composability and Customizability: Enables well-defined workflows and extensibility in a composable manner, allowing full customization in aspects such as model providers, logging, and orchestrators.
• Programmatic Control Flow: Keeps things simple as developers only need to write code without considering graphical nodes and edges. For branching logic, you can use if statements; for loops, use while loops.
• Human Input and Signals: Supports pausing workflows to receive external signals, such as human input, which can be used as tools that agents can call.

🚀 Quick Start

📦 Installation

To use MCP-Proxy, you first need to install the necessary dependencies:

pip install mcp-proxy[mpl,asyncio]

💻 Usage Examples

Here is a simple example demonstrating how to create a basic AI agent:

from mcp_proxy.agent import Agent
from mcp_proxy.llm import LLMOptions
import asyncio

# Create a configuration object
config = {
    "llm": {
        "provider": "anthropic",  # Supported: anthropic, openai, huggingface, etc.
        "model": "claude-2"
    },
    "logging": {
        "enabled": True,
        "level": "INFO",
        "output_file": "agent.log"
    }
}

# Create an LLM options object
llm_options = LLMOptions()

# Define the agent class
class MyAgent(Agent):
    async def main(self):
        # You can write any asynchronous logic here
        await self.call_tool("fetch_url", arguments={"url": "https://example.com"})

# Create and run the agent
async def run():
    agent = MyAgent(config=config, llm_options=llm_options)
    await agent.start()

if __name__ == "__main__":
    asyncio.run(run())

📚 Documentation

For more detailed information, please refer to the MCP-Proxy documentation.

🤝 Contribution

We welcome contributions in any form. Please check the CONTRIBUTING guide to learn how to get started.

Special Thanks

Several community members have made significant contributions to this project:

• Shaun Smith (@evalstate) - Led the team in numerous complex improvements, both in mcp-proxy and more broadly in the MCP ecosystem.
• Jerron Lim (@StreetLamb) - Contributed countless hours and excellent examples, and came up with great ideas for the project.
• Jason Summer (@jasonsum) - Identified many issues and adapted his Gmail MCP server for use with mcp-proxy.

🗺️ Roadmap

We plan to create a detailed roadmap based on your feedback. The current main priorities include:

• Persistent Execution - Allows workflows to pause/resume and serialize states so that they can be replayed or paused indefinitely. We are integrating Temporal to achieve this.
• Memory - Adds support for long-term memory.
• Streaming - Supports streaming listeners for iterative progress updates.
• MCP Extended Features - In addition to tool calls, supports the following:
  • Resources
  • Prompts
  • Notifications

❓ Frequently Asked Questions

What are the core advantages of using MCP-Proxy?

MCP-Proxy provides a simplified way to build AI agents by leveraging the capabilities provided by MCP servers.

MCP is a low-level protocol, and this framework handles tasks such as connecting to the server, interacting with large language models (LLMs), handling external signals (e.g., human input), and supporting persistent states, enabling you to focus on the core of business logic.

Core advantages:

• Interoperability: Ensures seamless integration of any tools provided by MCP servers.
• Composability and Customizability: Enables well-defined workflows and extensibility in a composable manner, allowing full customization in aspects such as model providers, logging, and orchestrators.
• Programmatic Control Flow: Keeps things simple as developers only need to write code without considering graphical nodes and edges. For branching logic, you can use if statements; for loops, use while loops.
• Human Input and Signals: Supports pausing workflows to receive external signals, such as human input, which can be used as tools that agents can call.

Do I need an MCP server to use MCP-Proxy?

Yes, MCP-Proxy requires an MCP server to provide tools and services. You can use an existing MCP server or create your own.

How can I extend the functionality of MCP-Proxy?

You can extend the functionality of MCP-Proxy by creating custom tools and commands. These can be defined in the Agent class and called through the call_tool method.

Which LLM providers does MCP-Proxy support?

Currently, the main supported providers include:

• Anthropic: Supports models such as Claude 2, Claude 3, etc.
• OpenAI: Supports models such as GPT-4, GPT-3.5, etc.
• Hugging Face: Supports various Transformer-based models.

More providers can be easily added by extending the framework.