The gap between reading a scientific paper and actually running its code is often bridged by hours of manual configuration and debugging. A research team from Stanford, led by Jiacheng Miao and James Zou, has introduced Paper2Agent, an automated framework designed to transform static research papers and their associated codebases into active, interactive AI agents.
Published in Nature on September 16, 2026, the system leverages the Model Context Protocol (MCP) to turn a paper's methods into an MCP server. This allows any compatible agent, such as Claude Code or Codex, to execute the research methods using natural language commands. This approach effectively creates a "virtual corresponding author" capable of reproducing results or running the original logic on entirely new datasets.
The Automation Pipeline
Paper2Agent operates via a central orchestrator that manages six specialized sub-agents to ensure the resulting tools are reliable. The process begins by locating and downloading the codebase, followed by the creation of an isolated virtual environment. A tutorial scanner then indexes usable guides, which a dedicated executor runs end-to-end to record reference outputs. Finally, a tool extractor converts these tutorials into parameterized MCP tools, which are then strictly validated.
The validation gate is rigorous: a tool only passes if the output files appear as expected and numerical results match within 3%. Visual figures are verified using perceptual hashes with a Hamming distance under 20. Tools that fail after six attempts are discarded, ensuring the final MCP server contains only verified functions.
Performance and Benchmarks
According to the official repository on GitHub, the system can automate complex biological tools with high efficiency. For AlphaGenome, Paper2Agent built 22 validated tools in approximately 45 minutes at a cost of 14 dollars. In benchmarks, it achieved 98.7% accuracy on tutorial-derived queries and a perfect 100% on novel queries, significantly outperforming standard "Claude + Repo" setups (78.7%) and Biomni (56%).
As reported by Runtimewire, the system is open-source and designed for scalability. Tests on 100 bioRxiv computational biology papers showed that 74 were successfully "agentified," with 593 out of 599 proposed tools passing validation. Aidailypost highlights that the framework also works outside biology, maintaining 98.1% accuracy on tasks involving models like SAM 2 and TabPFN.
Collaborative AI Science
The researchers demonstrated the power of these agents by connecting three distinct ones—AlphaGenome, an MPRA-coupled scCRISPRi screen agent, and a CD4+ T cell Perturb-seq dataset agent. Together, they identified GPR137 as the probable causal gene for a psoriasis locus, a hypothesis later supported by signature correlation under stimulation. This marks a shift toward multi-agent collaboration in scientific discovery, expanding on the automation of laboratory workflows previously explored by Claude Science.
The framework is available under an MIT license, with prebuilt servers for AlphaGenome, Scanpy, and TISSUE hosted on Hugging Face Spaces.
No comments yet. Be the first!