The End of Static Science
For centuries, the scientific paper has functioned as a static vessel – a document meant for human consumption. It is a record of discovery, but it is not, in itself, a tool. That paradigm is shifting. With the release of Paper2Agent, a framework published in Nature by researchers at Stanford Medicine, the research paper is being reimagined as active, deployable infrastructure. By converting manuscripts, code, and datasets into Model Context Protocol (MCP) servers, this framework effectively shortens the knowledge-to-agent pipeline to near zero.
The core innovation lies in the automation of agentification. Lead author Jiacheng Miao and co-author James Zou have developed a system that treats a research paper not as a final destination, but as a blueprint for an interactive AI agent. By leveraging MCP – which is rapidly becoming the industry standard for connecting large language models with external resources – Paper2Agent allows researchers to expose their workflows, data, and analytical tools as agent-native knowledge. The paper becomes a virtual corresponding author, capable of executing the very methods it describes.
Lean Deployment, Strong Results
The efficiency of this transformation is striking. The researchers demonstrated the framework across three distinct domains: AlphaGenome for genomics variant interpretation, Scanpy for single-cell transcriptomics, and TISSUE for spatial transcriptomics. The AlphaGenome agent, equipped with 22 MCP tools, was constructed in approximately 45 minutes at a cost of roughly $14. The Scanpy agent, with 7 tools, was similarly built in 45 minutes for $13.
Performance metrics suggest these agentified papers are highly capable research assistants. The AlphaGenome agent achieved 98.7% accuracy on tutorial queries and a perfect 100% on novel queries, significantly outperforming Claude with direct repository access (82.7%) and Biomni (37.3%). These figures indicate that when research is structured as an MCP-compliant agent, the model’s ability to navigate and apply specific scientific methodologies is vastly improved.
Agents Collaborating for Discovery
The true potential emerges when these agents interact. In a multi-agent collaboration demonstration, the AlphaGenome, Scanpy, and TISSUE agents were tasked with a complex biological problem. Working autonomously, they collaborated to identify and validate GPR137 as a causal gene for psoriasis at the rs887314 locus. This represents a fundamental shift: moving from static dissemination to interactive, multi-agent collaboration where research objects actively contribute to new discoveries.
The scalability was validated through large-scale evaluation. Of 100 computational biology papers, 74 were successfully agentified, with 593 of 599 proposed tools passing validation. Across 300 benchmark questions, the agents maintained 91.2% accuracy. The framework is open source on GitHub under an MIT license, with prebuilt MCP servers hosted on Hugging Face Spaces.
When papers become deployable MCP servers, the friction between reading a discovery and applying it vanishes. This is not just about faster research – it is about creating a cumulative, agent-native knowledge base where every new paper adds to the collective capability of the scientific community.