DEVELOPMENT OF A MULTI-AGENT FRAMEWORK BASED ON MODEL CONTEXT PROTOCOL

Abstract

The aim of this project report is to create a practical, extensible and modular AI assistant that can execute real-world productivity tasks, using the Development of a Multi-Agent Framework Based on Model Context Protocol. In a traditional chatbot setup, you’d have one big LLM responsible for all the interactions, resulting in context pollution, poor tool selection, limited scalability, and lack of maintainability when you integrate various domains like email, calendar, web search, weather, memory, file handling, and voice interaction. The proposed framework tackles these problems by means of a multi-agent architec ture, driven by a supervisor. Each user requests is analysed by a central supervisor agent and the most appropriate specialist agent is used. These dedicated agents include calen dar management, e-mail automation, web searching, weather retrieval, general assistance, memory operations and desktop file interaction. All agents employ domain-specific in structions and tools, making them more reliable and reducing the need to load unnecessary context. One of the key areas of this work is on the extension mechanism in the form of Model Context Protocol (MCP) for external capability servers.One important aspect of this work is on Model Context Protocol (MCP) as an extension mechanism for connecting external capability servers. The framework enables services like Gmail retrieval or access to files on the desktop to be made available via a uniform interface, allowing the framework to be extended without altering the basic flow. The system also operates an email subsystem that has a human in the loop, where emails are drafted, reviewed, re-written, if necessary, and only sent upon explicit consent. The framework has been designed to interact with text and voice using Streamlit or command-line interface. It also supports short-term memory via LangGraph checkpoint ing and long-term memory via SQLite FTS5. In conclusion, the project showcases the feasibility of assembling a practical and scal able framework of an AI assistant by integrating multiple agents, multimodal interaction, memory management, tool-augmented reasoning, and extensibility via MCP.