What is ScaleMCP
ScaleMCP is a tool selection method launched by PwC that dynamically equips large language model (LLM) Agents with Model Context Protocol (MCP) tools. Based on an automatic synchronization tool storage system and MCP servers, it solves the inefficiency and inconsistency problems caused by existing frameworks relying on manual updates of local tool libraries. The core of ScaleMCP is an automatic synchronization tool indexing pipeline, which ensures consistency between the tool storage and MCP servers based on CRUD operations (Create, Read, Update, Delete). ScaleMCP introduces the Tool Document Weighted Average (TDWA) embedding strategy, which can selectively emphasize key parts of tool documents, improving tool retrieval and Agent invocation performance.
Main Features of ScaleMCP
- Dynamic Tool Discovery and Equipping: LLM Agents dynamically discover and load the required MCP tools during multi-turn interactions without pre-configuration.
- Automatic Synchronization Tool Storage System: Maintains synchronization with MCP servers based on CRUD operations (Create, Read, Update, Delete), ensuring real-time updates and consistency of the tool storage system.
- Support for Multiple Retrieval and Embedding Models: Compatible with multiple LLM models, embedding models, and retriever types, with good scalability and flexibility.
- Improved Tool Invocation and Task Completion Rate: Improves the performance of LLM Agents in complex tasks, especially in multi-hop tool invocation scenarios.
Technical Principles of ScaleMCP
- Model Context Protocol（MCP）: MCP is a standardized protocol used to connect LLMs with external tools, data sources, and prompts, supporting developers in exposing tools or APIs based on MCP servers and simplifying the process for AI systems to access necessary data. ScaleMCP treats MCP servers as the single source of truth for tools and maintains synchronization with the tool storage system based on CRUD operations.
- Automatic Synchronization Tool Storage System: ScaleMCP is based on an automatic synchronization tool indexing pipeline that periodically retrieves the tool list from MCP servers and calculates the SHA-256 hash values of the tools. The newly retrieved tool hashes are compared with the existing hashes in the storage system, and if a mismatch is found, the tool index in the storage system is updated. ScaleMCP supports multiple storage systems, such as vector databases and graph databases, and selects the appropriate storage method based on the dependency relationships of the tools.
- Tool Document Weighted Average (TDWA) Embedding Strategy: Traditional tool document embedding methods (such as simple concatenation or averaging) cannot effectively distinguish the importance of different parts of tool documents. TDWA assigns weights to various parts of tool documents (such as tool name, description, parameters, synthetic questions, etc.) and calculates weighted average embedding vectors, more accurately representing the semantic information of tools.
- Dynamic Tool Retrieval and Invocation by LLM Agents: ScaleMCP provides LLM Agents with an MCP retrieval tool, through which the Agent retrieves relevant MCP tools based on keywords and loads them into the context. LLM Agents dynamically invoke multiple tools based on user queries, perform reasoning after obtaining tool responses, and ultimately generate answers. Agents can re-query the tool storage system during multi-turn interactions, dynamically adjusting tool selection and invocation strategies.
ScaleMCP Project Address
- arXiv Technical Paper：https://arxiv.org/pdf/2505.06416
Application Scenarios of ScaleMCP
- Financial Data Analysis: Dynamically loads financial APIs, updates data in real time, and quickly responds to complex financial queries.
- Intelligent Customer Service: Integrates enterprise APIs, invokes tools in real time, and provides precise customer support.
- Healthcare Consultation: Connects medical databases and appointment systems, supporting disease queries and appointment services.
- Educational Tutoring: Integrates school information systems, dynamically invokes tools, and queries course information and exam scores.
- Smart Office: Connects file management and scheduling systems, supporting file queries and meeting scheduling.