Stanford's Paper2Agent turns research papers into MCP agents

A Stanford team led by James Zou has published a framework in Nature that automatically converts research papers into interactive AI agents, and reports it successfully agentified 74 of 100 computational biology papers, hitting 91.2±1.6% accuracy across 300 benchmark questions.

The system, called Paper2Agent, wraps each paper's manuscript, supplementary materials, datasets and code into a Model Context Protocol server that a chat agent can call as tools. In the paper's headline demonstration, applying Paper2Agent to AlphaGenome produced 22 validated MCP tools in 45 minutes at a compute cost of about $14, and reached 100.0±0.0% accuracy on novel queries against 82.7±3.4% for a Claude model given the same repository.

"Paper2Agent addresses this challenge by converting a paper into an AI agent that functions as a virtual corresponding author, exposing its manuscript, supplementary materials, datasets, code and workflows as active, agent-native knowledge rather than static text," the authors write.

Beyond the single showcase, the team ran the framework across 100 computational biology papers and generated 599 proposed tools, of which 593 passed automated validation. Per-query inference cost fell to about $0.20 against $0.38 for the Claude + Repo baseline, and end-to-end runs came in 1.9-3.8× faster than the same baseline.

The authors also stage a small proof-of-concept of paper agents doing science together. Three Paper2Agent instances collaborated to nominate GPR137 as the probable causal gene at the psoriasis rs887314 locus, a result they validate through signature-correlation analysis rather than wet-lab work.

The framework's limits sit inside its own numbers. Only 74 of the 100 computational biology papers converted cleanly, and the abstract publishes no breakdown of the failure modes. Every paper agent depends on the underlying code being public and runnable, which excludes most published science outside computational fields. The paper has drawn more than 121,000 accesses since posting on 16 September 2026, and five of the AI researchers we track shared it in the first week.