jmiao24/Paper2Agent

Paper2Agent – Turning Research Papers into Interactive AI Agents

What it is – Paper2Agent is a Python‑based multi‑agent system that automates the conversion of a scientific paper (and its accompanying code repository, if any) into a Micro‑Component‑Program (MCP) server. The generated server exposes the paper’s methods as callable tools that can be accessed from AI coding assistants such as Claude Code, OpenAI Codex, or the Gemini CLI. In practice, a user asks their coding agent to install the Paper2Agent skill and then to agentify a paper; the skill spawns specialist sub‑agents that parse the paper, set up an isolated environment, wrap the original code, run verification tests, and finally bundle everything into a ZIP that can be run locally or hosted on Hugging Face.

Key capabilities

• One‑click conversion – Provide a URL or local path to a paper (and optionally its source repo) and the skill produces a ready‑to‑run MCP server.
• Parallel specialist agents – The coordinator launches separate agents for tasks such as literature parsing, dependency resolution, code testing, and verification, keeping the workflow reproducible.
• Verified, reproducible output – The generated ZIP contains a USAGE.md with the exact interpreter version, dependency list, and test results, ensuring the server works out‑of‑the‑box.
• Flexible deployment – The MCP can be run locally or uploaded as a Hugging Face Space; the README lists ready‑made hosted servers for AlphaGenome, Scanpy, and TISSUE.
• Secure handling of secrets – API keys are read from environment variables and never written into the generated code or notebooks.

Typical workflow

1. Install the skill – Copy the skills/paper2agent folder into the personal skill directory of your coding agent (paths differ for Claude Code vs. Codex).
2. Invoke the skill – In the agent’s chat, issue a prompt like:

   /paper2agent Convert https://github.com/scverse/scanpy into tested MCP tools in Scanpy_Agent.
   Focus on the "Preprocessing and clustering" tutorial.
   

3. Agent coordination – The skill spawns sub‑agents that:
   • download the paper and repo,
   • infer the required runtime (Python, R, GPU, etc.),
   • wrap the original functions as MCP endpoints,
   • run the paper’s own tests or supplied notebooks to verify correctness.
4. Receive a ZIP – The output contains the server code, a USAGE.md, and a USAGE.ipynb demonstrating the tools.
5. Connect to your coding assistant – Either unpack locally and run the server, or point the assistant to the hosted Hugging Face URL using the claude mcp add command.

Demo highlights

• AlphaGenome Agent – Accepts a gene‑variant query and returns the causal gene using the AlphaGenome MCP.
• TISSUE Agent – Provides uncertainty‑aware spatial transcriptomics predictions with confidence intervals.
• Scanpy Agent – Performs single‑cell preprocessing and clustering on a user‑supplied .h5ad file.

Installation checklist

• Python 3.x and Git installed on the host machine.
• A coding‑agent platform that supports skills and parallel sub‑agent spawning (Claude Code, Codex, or Gemini CLI).
• Optional: GPU or other native dependencies required by the target paper’s code.

Citation – If you use Paper2Agent in research, cite the Nature article:

@article{miao2026paper2agent,
  title={Reimagining research papers as interactive and reliable AI agents},
  author={Miao, Jiacheng and Davis, Joe R. and Zhang, Yaohui and Pritchard, Jonathan K. and Zou, James},
  journal={Nature},
  year={2026},
  doi={10.1038/s41586-026-11044-y},
  url={https://www.nature.com/articles/s41586-026-11044-y}
}

Bottom line – Paper2Agent lets you turn the methods described in a scientific publication into a plug‑and‑play AI tool, dramatically reducing the manual effort needed to reproduce and reuse cutting‑edge research.