Paper2Agent converts 74 of 100 bio papers into AI agents

Stanford researchers have built a system that turns a published research paper into an interactive AI agent, and they say it converted 74 of 100 computational biology papers into working tools at a median cost of $14 per paper.

The framework, called Paper2Agent, is described in Nature by Jiacheng Miao, James Zou and collaborators. It analyses a paper's manuscript and codebase, constructs a Model Context Protocol (MCP) server exposing the paper's data, code and workflows, then generates and runs its own tests to harden the result. That MCP can be wired into a chat agent such as Claude Code, so a reader asks scientific questions in prose and the paper's own tools execute in the background.

The authors describe the paper as functioning as "a virtual corresponding author," and argue that "papers become agent-native research objects rather than static documents." Across the 100-paper trial, the system generated 599 proposed tools and 593 of them passed automated validation. On a benchmark of 300 questions drawn from the papers' own tutorials, Paper2Agent scored 91.2 ± 1.6% accuracy, outperforming Claude Code operating on the same repositories directly. Median query latency dropped from 4.3 minutes to 1.6, at $0.20 per query versus $0.38.

A showcase agent built for the genomic variant tool AlphaGenome hit 98.7 ± 1.3% accuracy on tutorial-derived queries and 100% on novel ones, and the authors report it identified GPR137 as a causal gene for a psoriasis-associated variant.

Left unaddressed: how the MCP servers stay in sync as papers, datasets and code dependencies drift after publication, and what went wrong for the papers the system could not convert. Five researchers we track shared the paper on the day it dropped.