MCP tool discovery for autonomous LLM agents

MCP-Zero presents a novel approach to tool usage for large language model (LLM) agents, addressing the limitations of current systems that rely on static tool sets or retrieval-based selection. By reframing tool discovery as an active capability discovery problem, MCP-Zero empowers agents to autonomously decide when they need tools and to generate structured requests for them as tasks unfold. This method contrasts with traditional approaches that either overload agents with extensive tool schemas or assume static tool requirements, both of which hinder scalability and autonomy. MCP-Zero utilizes a two-stage hierarchical semantic routing process for efficient tool discovery, separating server selection from tool selection to maintain precision without overwhelming the agent. Its iterative, agent-driven process allows for continuous refinement and adaptation, supporting complex multi-step workflows and promoting a more sustainable and adaptable design pattern for MCP-based systems.