Paper2Agent Explained: How Stanford Turns Research Papers Into AI Agents, What It Costs and Where Humans Still Matter

A Nature paper describes Paper2Agent, which converts a research paper and its code into an MCP server that a chat agent can use. How it works, the AlphaGenome case built in about 45 minutes for US$14, accuracy results, limits, and how researchers can use it.

PUBLISHED // UPDATED
DURATION // 06 MIN READ
Printed scientific paper beside a laptop showing a simple chat window and a small stack of server rack stickers, cool daylight, no logos
FIG.01 // VISUAL DOCUMENTATION Printed scientific paper beside a laptop showing a simple chat window and a small stack of server rack stickers, cool daylight, no logos

Research papers describe methods, but using those methods usually means reading the paper, installing the code, working out the data formats and adapting everything to your own problem. A system called Paper2Agent, published in Nature on September 16, 2026, aims to shorten that process by turning a paper and its codebase into an AI agent that can answer questions and run the paper’s methods on new data. This article explains how it works, what the authors demonstrated, what it costs, and why human oversight remains essential.

What Paper2Agent does

The paper, “Reimagining research papers as interactive and reliable AI agents” by Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan Pritchard and James Zou of Stanford, describes an automated framework that converts a research paper into what the authors call a “virtual corresponding author”. It exposes the manuscript, supplementary materials, datasets, code and workflows as tools an AI agent can use.

The process works in three steps:

  1. Analysis. Multiple agents read the paper and its code repository to identify the key methods and workflows.
  2. Tool building. The system builds a Model Context Protocol (MCP) server, a standard way of exposing tools to AI assistants, with one tool per method or analysis step.
  3. Testing. It generates and runs tests to check that each tool works and reproduces the paper’s results, then refines the tools that fail.

The resulting “paper MCP” can be connected to a chat agent such as Claude Code. A researcher can then ask questions in natural language, and the agent calls the paper’s actual code to answer them. The paper also uses MCP prompts to encode multi-step workflows, so tools run in the right order without the user having to specify it.

As Nature’s news coverage summarises, the agent can answer questions about the paper, apply its methods to a fresh data set, and even collaborate with agents built from papers in other fields.

The case studies

AlphaGenome. The headline example wraps Google DeepMind’s AlphaGenome model, which predicts how DNA variants affect gene regulation. Paper2Agent generated 22 AlphaGenome MCP tools, all of which passed automated validation, in about 45 minutes at a cost of US$14 on a personal laptop, without human intervention. Across five runs, the resulting agent scored 98.7% accuracy on 15 queries derived from the tutorials and 100% on 15 new queries, compared with 82.7% and 78.7% for Claude given direct access to the code repository, and lower scores for the Biomni agent.

Scanpy. For the widely used single-cell analysis toolkit, Paper2Agent built seven tools for preprocessing and clustering in about 45 minutes for US$13, covering steps such as quality control, filtering and doublet detection.

TISSUE. A third agent handled spatial transcriptomics analysis using the TISSUE method.

Multi-agent discovery. In the most ambitious demonstration, agents built from several papers worked together to prioritise a causal gene for a psoriasis-associated genetic variant. The AlphaGenome agent identified GPR137 as the top affected gene at the locus in CD4+ T cells, and agents built from other studies were used to check that prediction.

The authors also report robustness tests: Paper2Agent correctly rejected all out-of-scope queries in a benchmark that paired papers with unrelated questions, and it recovered working servers after deliberately injected problems such as dependency errors, wrong file paths and deprecated APIs.

Why it matters

The low cost is the striking part. If building a reliable agent for a paper costs around the price of a lunch and takes under an hour, it becomes realistic to do it for many papers, not just famous ones. That could change how methods spread:

  • Faster adoption. Researchers in neighbouring fields could try a method without spending days on setup.
  • Better reproducibility checks. Because the tools are tested against the paper’s own results, failures surface quickly.
  • Cross-disciplinary work. Agents from different papers can be combined, as the psoriasis example shows.

The approach builds on earlier efforts such as executable papers, Papers with Code and container platforms like Binder and CodeOcean, but goes further by letting users interact through natural language.

What a session looks like

In practice, a researcher installs the generated paper MCP in a compatible assistant and asks questions in plain language, for example asking how a specific genetic variant affects gene expression in a given cell type. The assistant selects the relevant tool, runs the paper’s code with appropriate parameters and returns the result. Because each step is a tool call, the parameters and outputs can be logged and inspected later.

Where humans still matter

Automated agents for science raise familiar risks. A related review in environmental science, summarised by EurekAlert, makes the point clearly. Its authors see large language models helping with literature screening, mining relationships between pollutants and outcomes, and extracting quantitative data from papers. In one study they highlight, GPT-4 achieved 100% recall screening nearly 12,000 records and cut manual screening time by about half. But they stress that outputs can contain wrong numbers, mismatched units or unsupported relationships, and that “reliable applications need clear task definitions, structured constraints, traceable evidence, and human verification,” in the words of corresponding author Jing Guo of Nanjing University.

The same applies to Paper2Agent:

  • The agent is only as good as the code. If the original repository has bugs or undocumented assumptions, the agent inherits them.
  • Novel data can break assumptions. A method validated on one data type may give misleading results on another, even if the tool runs without errors.
  • Results need expert review. A confident answer from an agent is a starting point for a scientist to check, not a finding in itself.

How researchers can use it

  1. Try it on a method you already know. Compare the agent’s answers with your own results to understand its strengths and failure modes.
  2. Keep the tool calls. Record which tools the agent ran with which parameters so that analyses can be reproduced and audited.
  3. Validate on held-out cases before relying on outputs for new data.
  4. Consider publishing a paper MCP alongside your own code, so others can use your methods more easily.

For related research on AI in science, see our coverage of Claude and the Riemann hypothesis claims, testing vintage LLMs on relativity, and diffusion-based language models.

Our analysis

Paper2Agent is a practical step towards making research methods easier to use, and its cost of around US$14 per paper makes wide adoption plausible. The strongest results come from well-documented codebases such as AlphaGenome and Scanpy; less tidy research code will be harder to wrap. The right way to see these agents is as skilled assistants for applying a paper’s methods, with scientists still responsible for checking assumptions and interpreting results.