Stanford's Paper2Agent turns research papers into MCP agents

Paper2Agent, a framework from Stanford published in Nature on 16 September 2026, converts research papers into runnable AI agents by exposing their code, data and tutorials as Model Context Protocol servers that a language model can call in natural language. The authors (Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou) pitch it as a way to turn "passive artefacts into active systems that accelerate use and discovery."

The headline evidence is a large-scale run across 100 computational biology papers: 74 were successfully agentified, yielding 599 proposed tools of which 593 passed the framework's automated validation. On a 300-question benchmark drawn from those papers, the resulting agents averaged 91.2 ± 1.6% accuracy, against 82.7 ± 3.4% for a Claude plus repository baseline. For the AlphaGenome case study, the pipeline generated 22 tools in about 45 minutes at roughly $14 in model cost, and cut runtime 1.9× against direct repository access.

Five researchers on our radar circulated the paper the day it went up.

The system runs a six-stage pipeline: locate the codebase, build an isolated environment, scan for tutorials, execute them to capture reference outputs, extract parameterised MCP tools, then assemble them into a server. Validation is strict. Numerical outputs must match originals within a 3% tolerance and generated figures are compared by perceptual hash, and the extracted tools are locked to prevent later hallucination. In adversarial testing, the agents hit a 100% correct rejection rate on out-of-scope queries.

The paper does not soft-pedal its limits. 26 of the 100 biology papers stayed unagentifiable, blocked by missing code, unavailable data, environment failures, or scripts that would not generalise. The authors also flag that open-ended scientific reasoning "remains human-in-the-loop," with researchers still owning hypothesis selection and evidence evaluation, and that MCP servers built this way will need upkeep as their upstream dependencies drift.

In a multi-agent demonstration, three separate paper-agents (built from the AlphaGenome paper, an MPRA-coupled scCRISPRi paper, and a CD4+ T cell Perturb-seq paper) collaborated to prioritise GPR137 as a probable causal gene at the psoriasis-associated rs887314 locus, chaining computational predictions from one paper with experimental screens from the others. The authors frame the direction as "a collaborative ecosystem of AI co-scientists" where the paper is the interface.