MCP Agent Toolkit 🛠️

A modular, lightweight Python implementation of the Model Context Protocol (MCP) specification. Enables developers to rapidly build, test, and deploy custom tool-calling servers and sandboxed execution environments for AI agents (Cursor, Anthropic Claude, LangChain, AutoGen).

🌟 Key Features

• JSON-RPC 2.0 Compliant: Full implementation of MCP protocol methods (initialize, tools/list, tools/call).
• Zero Heavy Dependencies: Pure Python standard library implementation with zero external baggage.
• Sandboxed Filesystem Tool: Path traversal-proof file operations confined strictly to an isolated root directory.
• Agent Scratchpad & Memory Store: In-memory key-value state and audit trail logging for long-running workflows.
• Plug-and-Play Extensibility: Decorate or register any Python function as an MCP tool with schema validation in seconds.

🚀 Quickstart

1. Register Tools & Run Server

from mcp_agent_toolkit import MCPServer, MCPClient

# Initialize MCP Server
server = MCPServer(name="calculator-agent", version="1.0.0")

# Register custom tool
server.register_tool(
    name="multiply",
    description="Multiplies two numbers",
    parameters={
        "type": "object",
        "properties": {
            "x": {"type": "number"},
            "y": {"type": "number"}
        },
        "required": ["x", "y"]
    },
    handler=lambda x, y: x * y
)

# Connect Client
client = MCPClient(server)
result = client.execute_tool("multiply", x=6, y=7)
print(result["content"][0]["text"])  # Output: 42

2. Sandboxed File Operations

from mcp_agent_toolkit import SandboxedFilesystemTool

fs = SandboxedFilesystemTool(sandbox_root="./agent_workspace")
fs.write_file("notes.txt", "Autonomous agent task plan.")
content = fs.read_file("notes.txt")

🧪 Running Tests

python -m unittest discover tests

📄 License

Released under the MIT License. Built with clean architecture principles by Milind Chatrabhuji.