A Toolset-First Paradigm Based on Large Language Model and Agent via Model Context Protocol for Intelligent Computation

Abstract

Large language models (LLMs) and agents have lately propelled advances in artificial intelligence for science (AI4S). In this work, we propose a toolset-first paradigm for intelligent computation that decouples the tools from the agents via the model context protocol (MCP). We construct four MCP servers, HEA-MCP, VASP-MCP, LAMMPS-MCP, and OpenClaw-MCP. HEA-MCP provides theoretical calculations for high-entropy alloys (HEA). VASP-MCP and LAMMPS-MCP provide first-principles calculations and molecular dynamics simulations, respectively. OpenClaw-MCP contains an autonomous agent, OpenClaw, which renders the autonomous agent a callable tool. OpenClaw itself also supports the MCP to form a recursive agent architecture. By packaging the tools as MCP servers and connecting them to a general MCP-compatible agent, domain-specific capabilities can be obtained without building dedicated agents and can be reused across agentic frameworks.