# MCP-Zero: Proactive Toolchain Construction for LLM Agents from Scratch

**Authors:** Xiang Fei, Xiawu Zheng, Hao Feng  
**arXiv:** [2506.01056](https://arxiv.org/abs/2506.01056)  
**Date:** 2025-06-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2506.01056v1/

## Key points

- Authors Xiang Fei, Xiawu Zheng, and Hao Feng introduce the MCP-Zero framework for large language models (LLMs)
- Traditional approach of injecting tool schemas into prompts is costly and error-prone
- MCP-Zero enables LLMs to autonomously determine when and which external tools to retrieve by constructing a task-specific toolchain from scratch
- Framework consists of three key components: Proactive Tool Request, Hierarchical Vector Routing, and Iterative Proactive Invocation
- Evaluation using MCP-tools dataset shows that MCP-Zero reduces context overhead, accurately selects tools, decreases token consumption on APIbank while maintaining high accuracy levels, and supports multi-turn tool invocation with consistent accuracy
- Semantic grounding through sample demonstrations enhances model outputs by providing specific definitions for MCP servers and tools
- Demonstrated patch acts as a schema anchor for future work to enhance model understanding through grammar-based decoders
- Overall, MCP-Zero offers an innovative solution for proactive toolchain construction in LLM agents with improved efficiency in selecting external tools while maintaining high accuracy across tasks.

## Summary

In their paper "MCP-Zero: Proactive Toolchain Construction for LLM Agents from Scratch," authors Xiang Fei, Xiawu Zheng, and Hao Feng introduce a novel framework that addresses the challenges of integrating external tools into large language models (LLMs). The traditional approach of injecting numerous tool schemas into the prompt is costly and error-prone. To overcome this limitation, MCP-Zero enables LLMs to autonomously determine when and which external tools to retrieve by constructing a task-specific toolchain from scratch. The framework consists of three key components: Proactive Tool Request, Hierarchical Vector Routing, and Iterative Proactive Invocation. To evaluate their approach, the authors compiled MCP-tools dataset comprising 308 MCP servers and 2,797 tools extracted from the Model-Context-Protocol repository. Experimental results demonstrate that MCP-Zero effectively reduces context overhead and accurately selects tools from a large pool of candidates. It also significantly decreases token consumption on APIbank while maintaining high accuracy levels and supports multi-turn tool invocation with consistent accuracy across rounds. Additionally, the authors highlight the importance of semantic grounding provided by sample demonstrations in refining model outputs. By clarifying field meanings and providing specific definitions for MCP servers and tools, semantic matching becomes more precise. This demonstration patch acts as a schema anchor for future work to enhance model understanding through grammar-based decoders. Overall,<Organization> MCP-Zero presents an innovative solution for proactive toolchain construction in LLM agents. It offers improved efficiency in selecting external tools while maintaining high accuracy levels across various tasks.