统一MCP工具图是一个以研究为驱动的项目，将各种模型上下文协议（MCP）服务器的工具API汇聚和结构化为集中的Neo4j图数据库。该图作为一个智能基础设施层，使大型语言模型（LLMs）和代理型AI系统能够动态检索任何任务中最相关的工具——不会被冗余或令人困惑的选项淹没。

---

🚀 近期更新：该功能已集成到SmarterMCP预览模式中。

动态MCP服务器启动与最小工具上下文

🟢 动态MCP服务器编排

• 系统现在只启动满足特定用户查询所需的MCP服务器。五个常用的MCP服务器（包括动态工具检索MCP）默认保持温热状态;其他服务器按需启动，并在最后使用后保持活跃10分钟。
• 动态工具检索器MCP不仅返回工具元数据，还返回每个工具运行/连接MCP服务器所需的配置（自动从厂商的GitHub README获取）。

Unified MCP Tool Graph: A Intelligence Layer for Dynamic Tool Retrieval

Unified MCP Tool Graph is a research-driven project that aggregates and structures tool APIs from diverse Model Context Protocol (MCP) servers into a centralized Neo4j graph database. This graph functions as an intelligent infrastructure layer that enables large language models (LLMs) and agentic AI systems to dynamically retrieve the most relevant tools for any task — without being overwhelmed by redundant or confusing options.

---

🚀 Recent Updates: This has been integrated into SmarterMCP in the preview mode.

Dynamic MCP Server Spin-Up & Minimal Tool Context

🟢 Dynamic MCP Server Orchestration

• The system now spins up only the MCP servers required for a given user query. Five popular MCP servers (including the Dynamic Tool Retriever MCP) are kept warm by default; others are started on demand and kept alive for 10 minutes after last use.
• Dynamic Tool Retriever MCP returns not just tool metadata, but also the config needed to run/connect to the MCP server for each tool (fetched from the vendor’s GitHub README automatically).
• Automatic MCP Config Extraction: Uses the vendor’s GitHub repo to extract the MCP server config (from README) for each tool, so agents can spin up/connect to the right server on the fly.
• Error Handling: If config extraction fails, the system logs a warning and continues, ensuring robust tool retrieval.

🟢 Minimal Tool Context for LLMs/Agents

• Only the exact tools required for the user query are loaded into the agent’s context (not all tools from all MCP servers). This prevents LLM confusion and infinite tool loops.
• End-to-End Flow:
  1. User query is received.
  2. Dynamic Tool Retriever MCP queries the Neo4j graph and returns the top relevant tools plus their MCP server configs.
  3. The agent spins up/connects to only the required MCP servers (using the configs), and loads only the retrieved tools.
  4. The agent executes the workflow and returns the answer.

🟢 A2A and LangGraph Agent Support

• A2A Agent Example: See Example_Agents/A2A_DynamicToolAgent/ for a fully dynamic A2A agent that orchestrates MCP servers and tools per request.
• LangGraph Example: See Example_Agents/Langgraph/ for a LangGraph agent using the same dynamic, minimal-tool approach.

---

🔬 This repository focuses on the creation and evolution of the Unified Tool Graph Database. Chatbot-based integration (e.g., LangChain) is treated as a modular extension of this foundational layer.

📢 Support for Cline, IDE’s coming soon…

---

Research Problem