Stanford Paper2Agent Turns Research Into Reproducible AI Agents

Stanford University researchers have developed Paper2Agent, a novel system designed to transform scientific research papers and their associated code into interactive AI agents. This innovation aims to overcome the significant barrier of reproducing computational research, which often requires users to manually clone, install, configure, and debug codebases. The system was detailed in a paper published in Nature on September 16, 2026, and is led by Jiacheng Miao and James Zou. Paper2Agent converts a research paper and its codebase into a Model Context Protocol (MCP) server. This server can then be utilized by any MCP-compatible agent, such as Claude Code, to execute the paper's methodologies through natural language commands, effectively acting as a "virtual corresponding author." The system is designed for deployability, with its code released under an MIT license. It can be installed as a skill for AI agents like Claude Code or Codex. Pre-built servers for AlphaGenome, Scanpy, and TISSUE are available on Hugging Face Spaces, and a hosted version can be accessed at paper2agent.ai.

The Paper2Agent pipeline operates through a central orchestrator that dispatches specialized sub-agents across six distinct steps. Initially, the system locates and downloads the relevant codebase. An environment manager then establishes an isolated virtual environment for execution. A tutorial scanner identifies usable tutorials within the codebase, and a tutorial executor runs these tutorials end-to-end, recording the reference outputs. Subsequently, a tool extractor transforms these tutorials into parameterized MCP tools, which are then validated by a test verifier. The orchestrator assembles these validated tools into a single MCP server. The validation process is stringent: a tool is considered successful only if expected files are generated and numerical outputs match reference values within a 3% tolerance. Furthermore, figures must match references based on perceptual hashing, with a Hamming distance below 20. The verifier is allowed up to six attempts per function, and any tools that consistently fail are excluded from the final server. Each generated server comprises three key components: MCP tools, which encapsulate the paper's methods as executable functions; MCP resources, which store the manuscript, code links, datasets, and figures; and MCP prompts, which encode multi-step workflows, such as the correct order for data preprocessing in Scanpy. The research team utilized Claude Sonnet 4 for all Paper2Agent applications.