Stanford Publishes Paper2Agent to Turn Research Papers Into Working AI Tools

Stanford Medicine researchers have published Paper2Agent, a framework that turns a research paper’s manuscript, code, datasets and supplementary materials into an interactive AI agent. The resulting agents can answer questions, reproduce analyses, apply methods to new data and work with agents built from other papers.

The Nature paper, led by Jiacheng Miao and James Zou, targets a practical problem in computational research. Understanding a paper is often only the beginning: a researcher may still need to locate its code, install dependencies, configure an environment and determine how to use the method on a different dataset.

From publication to runnable method

Paper2Agent uses Model Context Protocol, or MCP, to package a study’s materials for AI systems. An MCP server holds executable tools based on the paper’s methods, resources such as its manuscript and datasets, and prompts that guide multi-step workflows. A connected chat agent can then invoke those tools through natural-language requests.

The conversion process tests its own work

The framework does more than summarize a PDF. Its worker agents identify a paper’s codebase, configure its software environment, extract core methods into tools, and run validation tests. Tools that repeatedly fail those checks are excluded from the final MCP server.

• Tools run core analytical methods in a preconfigured environment.
• Resources organize the manuscript, code, data, tables and figures.
• Prompts guide agents through workflows derived from the study.

A paired test produces an unvalidated lead

For a collaboration demonstration, the researchers paired an agent based on a genomic-mutation prediction paper with one based on an ADHD genome-wide association study. The mutation-prediction agent analyzed the ADHD dataset and prioritized variant rs1626703, proposing a possible effect on MPHOSPH9 splicing and expression in glutamatergic neurons.

Zou said an association near MPHOSPH9 and increased ADHD risk had not previously been reported. The authors say the proposed mechanism requires experimental validation. Yet papers may omit failed experiments and the judgment calls behind a setup, so authors can add that missing context in conversations with an agent.

The team has created more than 100 paper agents and has a longer-term goal of thousands or millions finding connections across the literature. That future depends on keeping those collaborations guided and monitored, while preserving credit for the original papers and human authors.