DMCP

About

DMCP addresses the significant challenge faced when integrating numerous Model Context Protocol (MCP) servers and their tools, which can lead to token explosion and LLM confusion due to the sheer volume of options. It solves this by implementing a two-process architecture that leverages semantic vector search. Instead of loading all tools upfront, DMCP allows an LLM to dynamically discover and retrieve only the most relevant tools in response to a specific query, thereby optimizing tool selection and reducing computational overhead.

Key Features

• Semantic tool discovery via specialized vector search
• Utilizes a fine-tuned ToolRet embedding model (e5-large-v2) for high accuracy
• Scalable vector store with Redis Stack (HNSW index) for rapid lookups
• Two-process architecture for efficient, dynamic tool retrieval
• Seamless integration with MCP clients through Streamable HTTP transport
• 2 GitHub stars

Use Cases

• Managing and integrating numerous MCP servers without overwhelming the LLM
• Dynamically providing relevant tools to AI systems based on user queries
• Enhancing large language model (LLM) agent efficiency in multi-tool environments