The paper "MCP-Zero: Active Tool Discovery for Autonomous LLM Agents" addresses the limitations of current Large Language Model (LLM) architectures that passively rely on pre-defined tool schemas injected directly into their prompts. This traditional approach causes massive context overhead and compromises genuine agent autonomy, while single-round retrieval methods fail to adapt to complex, multi-step tasks.

To solve this, the authors introduce MCP-Zero, a framework that shifts LLMs from passive tool consumers to autonomous capability architects by allowing them to discover and request tools dynamically. The framework relies on three core mechanisms:

• Active Tool Request: When an LLM identifies a capability gap, it autonomously generates a structured request specifying its exact tool and server requirements.
• Hierarchical Semantic Routing: A two-stage matching algorithm that first filters candidate servers based on platform requirements, then ranks specific tools based on semantic similarity to the agent's request.
• Iterative Capability Extension: Agents can progressively discover, evaluate, and integrate tools across multiple domains as a task evolves, allowing them to dynamically build toolchains and naturally self-correct if initial tools are insufficient.

To evaluate their framework, the authors constructed the MCP-tools dataset, a comprehensive collection of 308 servers and 2,797 tools gathered from the official Model Context Protocol (MCP) repository.

Experimental results demonstrate that MCP-Zero scales effectively with expanding tool ecosystems. It achieves a 98% reduction in token consumption compared to standard context-injection methods while maintaining or exceeding tool-selection accuracy, particularly in multi-turn conversations and large-scale environments.