Replicating research into executable knowledge

Abstract

This study presents an agentic pipeline that converts engineering research papers into callable Model Context Protocol (MCP) tools. Existing approaches such as Retrieval-Augmented Generation (RAG) can retrieve knowledge from scientific literature and generate responses for humans, but they do not address the translation of such knowledge into tools usable by a Large Language Model (LLM), leaving a gap between static literature and executable knowledge that LLM agents require. The proposed pipeline integrates PDF-to-Markdown conversion via MinerU, tool coordination through LangGraph, and MCP code synthesis using the OpenAI Codex CLI. The most suitable executable knowledge artifact is identified from the full document, which may be an algorithm, or a mathematical model. Implementation-relevant information is then extracted through the detailed-reading steps. The extracted information is consolidated into a task specification that drives the OpenAI Codex CLI to generate FastMCP-compatible MCP tools. The prototype is validated on papers containing specific algorithms and mathematical models. The proof-of- concept confirms the feasibility of converting unstructured scientific literature into usable MCP tools, and highlights the wider potential for extending callable tool libraries that enrich the LLM ecosystem.