A research paper works as an MCP server: Paper2Agent built agents from 74 of 100 papers

Paper2Agent was run on 100 biology papers: 74 turned into working agents on their own, without a human.

F. Fabulinnews writer

An agent based on the AlphaGenome paper was assembled without a human: 45 minutes and $14 of compute.

It contains 22 MCP tools. On tasks from the paper's tutorials, the agent answers with 98.7% accuracy, while on new queries not covered in the paper, it achieved 100%.

Behind this is a general approach: a paper stops being text to read. Paper2Agent parses the publication, its code, datasets, and described workflows, then turns them into tools that an agent calls by name.

Before and now. A method from a paper used to be reproduced manually: find the repository, set up the environment, understand the examples. Now the method arrives as a set of calls, and the agent applies it to your data itself.

The work was published in Nature on September 16, 2026, by authors from Stanford. In a run across 100 computational-biology papers, the agents proposed 599 tools, 593 of which passed automated validation. An open preprint from September 8, 2025 also describes the first scientific catch: an agent independently found a new splicing variant associated with ADHD risk.

How to install it. Paper2Agent lives as a skill inside an agent: `git clone https://github.com/jmiao24/Paper2Agent.git`, then copy the skill folder to `$HOME/.claude/skills/paper2agent` (for Codex, the path is `$HOME/.agents/skills/paper2agent`). Next comes a slash command such as `/paper2agent Convert https://github.com/sunericd/TISSUE into tested MCP tools in TISSUE_Agent.` Three hosts are supported: Claude Code, Codex, and Gemini CLI; the system needs Python and Git.

Three agents are already ready. AlphaGenome, Scanpy, and TISSUE are hosted on Hugging Face Spaces and connect with a single command: `claude mcp add --transport http <name> <address>`, while the list is checked through `claude mcp list`. The Scanpy agent for single-cell data analysis took 45 minutes and $13 to assemble; it has 7 tools.

Across 300 tutorial questions, MCP-wrapped tools delivered 91.2% accuracy versus 86.3% for the model itself. A query cost $0.20 and took 1.6 minutes, versus $0.38 and 4.3 minutes.

Automation could not handle 26 of the 100 papers: they still need human hands.

A research paper works as an MCP server: Paper2Agent built agents from 74 of 100 papers

Paper2Agent was run on 100 biology papers: 74 turned into working agents on their own, without a human.

F. Fabulinnews writer

An agent based on the AlphaGenome paper was assembled without a human: 45 minutes and $14 of compute.

It contains 22 MCP tools. On tasks from the paper's tutorials, the agent answers with 98.7% accuracy, while on new queries not covered in the paper, it achieved 100%.

Across 300 tutorial questions, MCP-wrapped tools delivered 91.2% accuracy versus 86.3% for the model itself. A query cost $0.20 and took 1.6 minutes, versus $0.38 and 4.3 minutes.

Automation could not handle 26 of the 100 papers: they still need human hands.