Stanford Team Releases Paper2Agent for Research Papers

Updated 19 Sep 2026 · Reviewed by the Zekai editorial team

Computational research papers typically require readers to clone, install, configure, and debug code before seeing results. To address this hurdle, a Stanford team led by Jiacheng Miao and James Zou published Paper2Agent on September 16, 2026. The system converts an academic paper and its associated codebase into a Model Context Protocol server, acting as a virtual corresponding author for readers.

The MIT-licensed software installs as a skill for Claude Code or Codex, with prebuilt AlphaGenome, Scanpy, and TISSUE servers hosted on Hugging Face Spaces. Operating on Claude Code's agent SDK, a central orchestrator deploys specialized sub-agents through a six-step workflow to wrap methods, hold manuscript resources, and encode multi-step procedures.

Validation relies on a strict verification gate. A tool only passes if expected files appear, numerical outputs match within a three percent margin, and figures achieve a Hamming distance under twenty using perceptual hashing. The verifier permits up to six attempts per function, dropping any tools that continue to fail.

For AlphaGenome, Paper2Agent built twenty-two tools in roughly forty-five minutes at a cost of US $42, with every tool passing validation without human intervention. In five runs graded by two human experts with high agreement, median runtimes fell significantly compared to baseline setups using Claude with repository access or Biomni.

The Scanpy agent generated seven validated tools in about forty-five minutes for US $13, successfully matching human researchers across four public datasets for cell counts, gene counts, and top marker genes. A TISSUE agent similarly reproduced human results on spatial transcriptomics data without manual code cleanup.

Why it matters for AI for Students, Study & Research

For students and researchers bogged down by the friction of cloning, installing, and debugging code from academic PDFs, Paper2Agent offers a streamlined route to executing study methods. By turning repositories into interactive servers accessible via natural language, it lets you test results and replicate workflows faster during literature reviews.

What to do about it

• Explore the hosted version of the tool at paper2agent.ai
• Install the MIT-licensed code as a skill for Claude Code or Codex
• Test prebuilt AlphaGenome, Scanpy, and TISSUE servers on Hugging Face Spaces

Tools mentioned

• GitHub
• Claude
• Claude Code

Source: Stanford Researchers Release Paper2Agent: Turning Research Papers Into AI Agents That Reproduce Results and Run on New Data – MarkTechPost (marktechpost.com)