Stanford Paper2Agent Turns Research Papers Into MIT-Licensed MCP Agents

Stanford researchers published Paper2Agent in Nature on September 16, 2026, describing an automated multi-agent pipeline that turns a research paper and its public codebase into a Model Context Protocol (MCP) server that chat agents can call in natural language. The accompanying open-source repository is licensed under MIT and ships remote MCP demos for AlphaGenome, Scanpy, and TISSUE on Hugging Face Spaces.

Paper2Agent’s pipeline extracts tools from tutorials and methods, configures a reproducible environment, then iteratively generates and runs tests so each tool matches reference outputs before the MCP is assembled. Once hosted remotely, any MCP-capable agent can invoke the paper’s methods without cloning or hand-configuring the original stack.

What shipped

The Nature paper and arXiv HTML report case studies where Paper2Agent built an AlphaGenome agent with 22 tools, a TISSUE agent for uncertainty-aware spatial transcriptomics, and a Scanpy agent focused on preprocessing and clustering. Code availability points to github.com/jmiao24/Paper2Agent plus hosted MCP Spaces under the Paper2Agent organization. The authors frame the release as shifting papers from passive artifacts into interactive co-scientists, including a demo that links AlphaGenome tools with ADHD GWAS data for autonomous hypothesis exploration.

Treat tutorial and novel-query accuracy figures in the paper as author-reported evaluations. This brief does not invent independent benchmarks beyond those primaries.

Primary sources are the September 16 Nature article, the arXiv HTML manuscript, the MIT-licensed GitHub repository, and the public Hugging Face MCP Spaces.