Reading a computational paper is easy. Running its code is a nightmare. You clone repos, fight dependency conflicts, and debug environment errors before you even see a result. A Stanford team led by Jiacheng Miao and James Zou has built a bridge over this gap. Their tool, Paper2Agent, published in Nature on September 16, 2026, automatically converts academic papers and their codebases into Model Context Protocol (MCP) servers. This lets any compatible AI agent execute the paper’s methods through simple natural language prompts.
The system acts as a virtual corresponding author. It doesn’t just read the text; it builds a working software environment. An orchestrator dispatches sub-agents to download code, build isolated virtual environments, and run tutorials end-to-end. The validation gate is strict. A tool only passes if output files appear and numbers match reference data within 3%. Figures must also align via perceptual hash checks. If a function fails six times, it gets cut. This rigor ensures the resulting MCP server offers reliable, executable tools rather than broken scripts.
The efficiency gains are stark. Creating the AlphaGenome agent took about 45 minutes and cost just $14 using Claude Sonnet 4. All 22 generated tools passed validation without human help. In benchmark tests across 100 bioRxiv papers, Paper2Agent successfully agentified 74 of them. It answered queries with 91.2% accuracy, beating standard Claude Code setups by over 10 percentage points. Costs dropped to $0.20 per query, down from $0.38, while runtime slashed from 4.3 minutes to 1.6 minutes.
This isn’t just about speed. It enables new science. When the team connected three different paper agents, they identified GPR137 as a likely causal gene for psoriasis, a finding the original papers had missed. The AI co-scientist proposed validation strategies that matched experimental signatures. With prebuilt servers now live on Hugging Face Spaces and an MIT license, the barrier to reproducing complex computational biology is collapsing. The question is no longer whether you can run the code, but whether you trust the agent to interpret it correctly.
