Paper2Agent in Nature: Research Papers Become Working Tools

A method published in Nature on September 16, 2026 takes a research paper and its code repository and produces a set of tested tools that an AI agent can run on request. The Stanford authors call it Paper2Agent. Instead of cloning a repository and fighting its dependencies, a user asks the agent in plain language to apply the paper's method to a new dataset.

This post is for technical founders, analysts and research leads who depend on methods that arrive as a paper plus code, and for builders who want the pattern for any codebase. It draws on the Nature article by Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan Pritchard and James Zou, received by the journal on October 13, 2025 and accepted August 14, 2026. An arXiv preprint dates from 2025. All figures below are the authors' own.

A paper describing a computational method is, in the authors' word, passive. To use it you find the repository, install dependencies, set up the environment and work out what inputs the code expects. Their example is AlphaGenome, a model that predicts the effects of DNA variants. Using it in code means installing an environment, creating API clients, building input objects and learning the parameter hierarchy, which is a real barrier for a biologist who only wants the answer.

Paper2Agent's answer is to package the method behind the Model Context Protocol, or MCP, the standard interface through which AI agents call external tools. Once a paper is an MCP server, the request becomes a sentence. The article's example is "Apply the method in this paper to the newly generated dataset".

The system is itself a group of agents, built on Claude Code. The authors state they used Claude Sonnet 4 for all applications in the paper, so nothing here says which model works best. The steps:

1. Find the code. Identify the repository that belongs to the paper.
2. Build the environment. An environment agent installs what the code needs so it actually runs.
3. Extract tools. An extraction agent turns the paper's core methods, usually shown in its tutorials, into callable functions with documented parameters.
4. Test against known outputs. A testing agent runs each tool on the paper's example data and compares the result with the reference. A test passes when expected files appear, numbers fall within tolerance and figures match. Code and environment are revised until it does, and tools that keep failing are excluded.
5. Package and host. Validated tools and their environment become one MCP server, which can be hosted remotely. The authors name Hugging Face Spaces as an example host.

The server carries three kinds of content, which map onto standard MCP concepts.

The locking step is the part to notice. Once a tool passes, its code is fixed. The agent calls it and does not rewrite it on each request. The authors present this as the defence against what they call "code hallucination", where an agent writes plausible new code that quietly computes the wrong thing.

Every number in this table is from the Nature article and was produced by its authors. Accuracy figures are printed as the article gives them, a mean plus or minus its error. The AlphaGenome benchmark used five runs graded by two human experts against set rubrics. "Claude + Repo" is the authors' baseline: Claude Code given a local copy of the repository.

Two details stand out. The wrapped tools on an older model beat the raw-repository approach on a newer one, 91.2% against 86.3%. That suggests, in this test, that preparing the code helped more than upgrading the model. And for the Scanpy case the authors report that the agent's output on four public datasets matched what human researchers produced on the same data: the same cell and gene counts after quality control and equivalent top marker genes.

For 26 papers that were about data or findings, with no method to run, only the resource layer was built. The authors report 89.0% accuracy on 100 synthesis questions against 82.0% for a browser-based baseline, at 34 times lower cost. We make no claim about the scientific findings in any case study.

Nothing in the pipeline is specific to science. Most companies have a version of the same problem: an analysis repository one person understands. A pricing model, a churn score, a forecasting notebook, a data-cleaning script with a README from 2023. The knowledge is in the code, and using it means finding its author.

The Paper2Agent recipe carries over if the repository has what the papers had: a worked example with known outputs. That example is the test. Wrap each function as a tool, run it on the example, compare with the saved result, and only expose tools that match. An analyst can then ask an agent to run the churn model on this month's export without opening the code.

If you want to build one by hand first, our tutorial on building an MCP server in TypeScript covers the server side. For judging how strong a paper's evidence is before you wrap its method, see the reference on types of proof in AI research. The project's GitHub repository packages the method as a skill for coding agents, and its README names Claude Code and Codex as supported.

Twenty-six of the 100 biology papers could not be converted. The article lists the causes: missing executable code, missing data or model files, environment and dependency failures, and scripts too specific to generalise. The authors suggest that how easily a paper converts may itself be a practical measure of its reproducibility. The same is true of a company repository.

They add four cautions that apply outside science too:

• Open-ended reasoning stays with people. The agent can run analyses at scale, and researchers remain responsible for choosing directions and judging evidence.
• On open-ended questions a benchmark that scores agreement with one reference answer measures faithful execution, not whether the analysis is valid.
• Wrapped tools need maintenance as the underlying code and dependencies change.
• Exposing code as callable tools raises security, intellectual property and attribution questions.

The last point matters more for a business than for a lab. A tool server that can run your pricing model is a system to put behind authentication and access control. Our AI transformation practice builds internal tool servers with those controls in place.