A research paper that can answer questions, run its own methods, and collaborate with other agents is no longer a thought experiment. Stanford Medicine researchers have built it. According to Forkast, a team led by Jiacheng Miao and James Zou published Paper2Agent in Nature, a framework that converts scientific manuscripts, code, and datasets into Model Context Protocol servers. The result is an AI agent that does not just describe a method. It executes it.
The timing matters. MCP has moved fast from a niche Anthropic proposal to something close to an industry standard for connecting large language models with external tools and data sources. OpenAI, Google DeepMind, and most major agent frameworks have adopted or acknowledged it. Paper2Agent is betting that MCP becomes the connective tissue for scientific knowledge, not just software products. That’s a reasonable bet right now.
The mechanics are straightforward. The framework reads a paper’s manuscript, its associated code repository, and any linked datasets, then auto-generates MCP-compliant tools that expose the paper’s core workflows to any compatible LLM. The paper stops being a static document and becomes what the researchers call a virtual corresponding author. You can query it, run its analyses, and get outputs grounded in the actual methodology it describes.
The accuracy numbers are the most compelling part of this work. The AlphaGenome agent, built with 22 MCP tools in roughly 45 minutes at a cost of about $14, hit 98.7% accuracy on tutorial queries and 100% on novel queries. Compare that to Claude with direct repository access at 82.7%, or Biomni at 37.3%. Structuring research as an MCP-compliant agent clearly helps a model apply specific scientific methods more reliably than just pointing it at a GitHub repo and hoping for the best.
The multi-agent demonstration pushes the concept further. AlphaGenome, Scanpy, and TISSUE agents, covering genomics, single-cell transcriptomics, and spatial transcriptomics respectively, were given a complex biological problem and left to collaborate. They independently identified and validated GPR137 as a causal gene for psoriasis at the rs887314 locus. That kind of cross-domain autonomous reasoning is exactly what AI-assisted research has been promising for years without quite delivering.
Scale validation adds credibility. Of 100 computational biology papers tested, 74 were successfully agentified. Across 300 benchmark questions, the agents held 91.2% accuracy. The framework is open source under an MIT license, with prebuilt servers on Hugging Face Spaces.
For developers building research tools, this is worth watching closely. The core value proposition is simple: the gap between reading a discovery and actually using it collapses. Every agentified paper becomes part of a growing, queryable knowledge base. And as more papers get converted, the compounding effect on scientific productivity could be significant. The question now is whether the research community adopts the tooling or waits for institutions to mandate it.
