Nature study turns research papers into tested MCP agents

A peer-reviewed Nature study introduces Paper2Agent, an automated framework for turning a computational research paper into an executable agent interface. The system ingests the manuscript, supplementary material, datasets and associated code, then uses specialised agents to configure the software environment, extract core methods as tools, build a Model Context Protocol server and test those tools against the paper's reference outputs.

The validation step is central to the design. Generated tools are accepted only when expected files are produced, numerical results stay within defined tolerances and reproduced figures match the reference. Tools that repeatedly fail are excluded. Each accepted tool also retains a reference to the original paper code, giving the downstream agent a traceable implementation rather than asking a language model to improvise scientific code from prose.

The authors tested the approach on methods including AlphaGenome, Scanpy and TISSUE. For AlphaGenome, Paper2Agent generated 22 validated MCP tools in about 45 minutes on a personal laptop at a reported one-time model cost of $14. The resulting agent could score genetic variants, explore tissue-specific effects and generate visualisations through natural-language requests.

The paper also combines multiple paper agents in a psoriasis case study, using separate tools to prioritise a candidate causal gene. Across the reported demonstrations, the agents reproduced analyses from the source papers and then applied the same methods to new user queries. Claude Sonnet 4 and Claude Code were used in the study's implementations, while the generated MCP servers are designed to be callable by other compatible agents.

Computational papers often publish enough information to reproduce a method in principle while still leaving a reader to locate repositories, install dependencies, understand APIs and reconstruct analysis steps. Paper2Agent tries to package those moving parts as a reusable, machine-callable research object. That makes the contribution closer to an executable interface around a paper than to ordinary retrieval-augmented question answering over its text.

The use of MCP also lets several paper-specific toolsets coexist behind one agent. In principle, a researcher could combine validated methods from different publications without manually wiring every codebase together. The stronger claim is therefore about reducing friction around computational method reuse and reproducibility, rather than replacing scientific judgement or the original publication.

The evidence comes from a small set of computational biology case studies whose code could be packaged and tested automatically. It does not show that arbitrary experimental, theoretical or poorly documented papers can be converted reliably, and it does not remove errors already present in the source code, data or publication. The approach also depends on software environments and external services remaining available.

The paper is a peer-reviewed version of record, but the broader proposition still needs testing across more disciplines, older codebases and methods with ambiguous or partially manual workflows. The useful next evidence will be independent attempts to generate paper agents from unfamiliar publications and checks of whether researchers can reproduce, audit and extend the resulting analyses without hidden implementation drift.