Paper2Agent framework converts research papers into runnable AI agents via MCP

A new paper published in Nature introduces Paper2Agent, an automated framework that converts research papers into runnable AI agents by packaging their code, data, and methods into a Model Context Protocol (MCP) server. This allows chat assistants to directly call the paper's computational tools rather than just reading its text.

Performance and Cost

In a case study using Google DeepMind's AlphaGenome, the framework generated 22 validated tools in approximately 45 minutes at a compute cost of roughly $14. It achieved 98.7 ± 1.3% accuracy on tutorial-derived queries, significantly outperforming a baseline that provided direct repository access to Claude (82.7%).

Broader Applicability

On a wider set of 100 computational biology papers, Paper2Agent successfully created usable agents for 74 of them, scoring 91.2 ± 1.6% on 300 benchmark questions. The authors note that about a quarter of the tested papers failed due to missing code, environment issues, or non-generalizable scripts.

Implications for Research

The authors argue that this approach transforms research from passive artifacts into active systems, potentially accelerating adoption and discovery. One case study claimed an assembled AI co-scientist identified a new splicing variant associated with ADHD risk. The paper suggests journals should add an "agent availability" clause to existing data and code requirements.