Content
# MCP-AG2 Integration Example
This project demonstrates the integration of the [Model Context Protocol (MCP)](https://modelcontextprotocol.io/) with [AutoGen (AG2)](https://github.com/ag2ai/ag2), showcasing a powerful pattern for building modular, tool-enabled AI agents.
## Overview
The example implements three key components:
1. **MCP Server**: A process that exposes resources and tools following the MCP specification
2. **MCPAssistantAgent**: An AutoGen AssistantAgent extension that implements the MCP client interface
3. **Example Script**: Demonstrates an MCP-enabled agent using LLM capabilities with MCP resources/tools
## Setup
### Prerequisites
1. Install `uv` package manager:
```bash
# macOS
brew install uv
# Other platforms
curl -LsSf https://astral.sh/uv/install.sh | sh
```
### Installation
```bash
# Clone the repository
git clone https://github.com/jtanningbed/mcp-ag2-example
cd mcp-ag2-example
# Install dependencies
uv sync
# Run the example
uv run example.py
```
## Key Benefits
This integration pattern offers several advantages over traditional tool/function calling implementations:
### 1. Protocol-Level Interface Abstraction
- The MCP client interface itself (`read_resource`, `call_tool`, etc.) is exposed to the LLM agent through AG2's tool registration
- Rather than registering individual tools directly with AG2, we register only the core MCP interface methods
- All specific tools (e.g., `write_file`) are proxied through the `call_tool` interface to the MCP server
- This means tools defined on the MCP server don't need any format conversion for different LLMs - they remain in Anthropic schema format
### 2. Dynamic Tool Discovery and Model Agnosticism
- Agents use `list_tools` to discover available server tools
- While the MCP server defines tools using Anthropic's schema format, the LLM never sees these directly
- The LLM only needs to understand how to use `call_tool` with a name and arguments
- No need to convert tool schemas between different LLM formats since they're abstracted behind the MCP interface
### 3. Clean Separation of Concerns
- MCP server handles:
- Tool implementation details
- Tool schema definitions (in Anthropic format)
- Resource management
- Agent only handles:
- Understanding the core MCP interface methods
- Using `call_tool` to proxy specific tool requests to the server
- LLM integration layer only handles:
- Registering MCP interface methods as tools (e.g., `call_tool` in OpenAI format)
- Routing tool calls through the MCP client
### AutoGen-Specific Benefits
When compared to traditional AutoGen tool implementations:
1. **Simplified Tool Integration**
- No need to define tool schemas in multiple formats
- Tools are defined once on the MCP server in Anthropic format
- AG2 only needs the `call_tool` interface registered
2. **Enhanced Modularity**
- MCP servers can be used by any MCP-compatible client
- Tools and resources are completely decoupled from agent implementation
- New tools can be added to the server without any agent changes
## Architecture Overview
```mermaid
graph LR
A[LLM] -->|Uses| B[AG2 Tool Interface]
B -->|Registered| C[MCP Interface Methods]
C -->|Proxy| D[MCP Client]
D -->|Protocol| E[MCP Server]
E -->|Implements| F[Tools & Resources]
```
MCP Store