Paper2Agent turns research papers into interactive AI agents

Nature has published Paper2Agent, a framework that turns research papers and their associated code, data and supplementary material into interactive AI agents. The authors also made the Paper2Agent repository publicly available under an MIT license, giving researchers a way to build paper-specific tools that respond to natural-language requests.

The release makes a paper’s methods executable through a paper-specific Model Context Protocol server. In the authors’ end-to-end test of 100 computational-biology papers, Paper2Agent converted 74 and generated 599 proposed tools. Of those, 593 passed the system’s automated validation. The study team produced the results, which have not been independently replicated.

Paper2Agent has two main layers. Paper2MCP analyzes a manuscript and codebase, prepares the software environment and builds the MCP server. An agent layer connects that server to a conversational AI system. The server can expose executable tools for a paper’s methods, resources such as manuscript text and supplementary files, and prompts that encode multi-step workflows.

The conversion pipeline scans and runs tutorials, captures reference outputs and turns reusable steps into MCP tools. It creates a test for each function, checks the generated tool against the reference output and excludes tools that repeatedly fail. The system then assembles the validated modules into an MCP server.

That validation is designed to catch errors in generated tools, not to establish that a paper’s scientific findings are correct. In the implementation described in the paper, tests check whether expected files are produced, compare numerical results within a 3% floating-point tolerance and compare figures using perceptual hashes with a Hamming distance below 20.

The researchers implemented Paper2Agent with Claude Code and used Claude Sonnet 4 for the applications reported in the paper. They say the generated MCP servers can connect to other compatible agents, but the available evidence does not establish equivalent reliability across those alternatives.

In one case study, the system generated 22 AlphaGenome tools in about 45 minutes at a reported cost of $14 on a personal laptop; all 22 passed the authors’ automated checks. Across five runs, the study team reported 98.7% accuracy on 15 tutorial-derived queries, 100% on 15 novel queries and 82.7% on 30 open-ended queries.

At larger scale, the authors reported 91.2% accuracy on 300 tutorial-derived questions using Sonnet 4. Their direct-repository comparison scored 80.3% with Sonnet 4 and 86.3% with Sonnet 4.6. For 42 execution tasks drawn from 10 non-biology computational papers, the team reported 98.1% accuracy. A separate resource-only evaluation across 26 data- and discovery-focused papers scored 89% on 100 synthesis questions.

The paper attributes the 26 unsuccessful computational-biology conversions to missing executable code, data or model artifacts, environment and dependency failures, and scripts that could not be generalized. The public repository lists prerequisites that include a coding-agent host with skills, shell access and parallel subagent support, plus Python, Git and any runtimes required by the target research code. The repository page does not show a formal versioned release or establish the code’s first-publication date.