Published in Nature on September 16, the framework converts an article and its code into an MCP server, turning the publication into an executable tool rather than a fixed text.
Paper2Agent, presented by James Zou's team at Stanford University and published in Nature on September 16, 2026, shares a common observation: to reuse a method described in an article, one must read the text, understand the code, adapt it to its own data. These steps hinder the actual dissemination of the work.
The framework automates the conversion of an article and its code repository to a Model Context Protocol server. Any MCP-compatible agent — Claude Code, for example — can then execute the article's methods in natural language, reproduce its results or apply them to a new dataset. The article ceases to be a passive artifact: manuscript, additional material, data and workflow become questionable, the agent holding the role of a corresponding author permanently available.
The Stanford Medicine researchers went further by engaging several agents from different articles, combining their respective methods to produce analyses that neither would have conducted alone.
Read the full article on the original source
Go to the source →