Paper2Agent: Reimagining Papers As AI Agents

📖 Overview

Paper2Agent is a multi-agent AI system that automatically transforms research papers into interactive AI agents with minimal human input. Explore demos of Paper2Agent-generated agents, or try it yourself at paper2agent.ai.

Paper2Agent coordinates parallel specialist agents to turn scientific papers into reliable MCP servers or skills.

🚀 Quick Start

Basic Usage

The simplest way to use Paper2Agent is to ask your coding agent (Claude Code, Codex, etc.) to install the skill, then agentify a paper alongside its code repository.

1. Ask your coding agent to install Paper2Agent:

Read https://github.com/jmiao24/Paper2Agent and install the paper2agent skill
from skills/paper2agent for this coding agent. Include the entire skill folder
with its references, scripts, and agents.

If the skill does not appear after installation, restart your coding agent. For manual installation, see Installation & Setup.

2. Ask it to agentify a paper alongside its code repository:

Use the paper2agent skill to agentify this paper alongside its code repository.
Paper: <PAPER_URL_OR_LOCAL_PDF>
Code repository: <GITHUB_URL_OR_LOCAL_PATH>
Output directory: <PROJECT_DIR>

You can also invoke the skill explicitly:

Claude Code:

/paper2agent Agentify <PAPER_URL_OR_LOCAL_PDF> alongside its code repository <GITHUB_URL> in <PROJECT_DIR>.

Codex:

$paper2agent Agentify <PAPER_URL_OR_LOCAL_PDF> alongside its code repository <GITHUB_URL> in <PROJECT_DIR>.

The skill selects useful operations from the repository's APIs, tutorials, examples, and tests. A completed conversion delivers dist/<repo-name>-mcp.zip with installation and usage instructions. Processing time and cost depend on the selected scope, dependencies, hardware, and coding-agent model.

Advanced Usage

Targeted Tasks or Tutorials

Specify the scientific tasks, tutorial title, or source URL to focus on:

Use the paper2agent skill to convert <GITHUB_URL> into MCP tools in <PROJECT_DIR>.
Focus on <TASKS, TUTORIAL_TITLE, or SOURCE_URL>.

Repository with API Key

Make credentials available through your host's secret mechanism or process environment, then tell the agent the variable name:

Use the paper2agent skill to convert <GITHUB_URL> into MCP tools in <PROJECT_DIR>.
Read the required API key from the environment variable <VARIABLE_NAME>.

Credentials stay outside generated code, notebooks, reports, and the delivered ZIP.

Inputs

Request optional extensions, such as user-query evaluation or remote deployment, when needed.

Examples

The examples below use Claude Code's /paper2agent invocation. In Codex, replace it with $paper2agent.

TISSUE Agent

Create an AI agent from the TISSUE research paper codebase for uncertainty-calibrated single-cell spatial transcriptomics analysis:

/paper2agent Convert https://github.com/sunericd/TISSUE into tested MCP tools in TISSUE_Agent.

Scanpy Agent for Preprocessing and Clustering

Create an AI agent from the Scanpy research paper codebase for single-cell analysis preprocessing and clustering:

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

You can also provide a tutorial URL:

/paper2agent Convert https://github.com/scverse/scanpy into tested MCP tools in Scanpy_Agent.
Focus on https://github.com/scverse/scanpy/blob/main/docs/tutorials/basics/clustering.ipynb.

AlphaGenome Agent

Create an AI agent from the AlphaGenome research paper codebase for genomic data interpretation:

/paper2agent Convert https://github.com/google-deepmind/alphagenome into tested MCP tools in AlphaGenome_Agent.
Read the API key from the environment variable ALPHAGENOME_API_KEY.

⚙️ Installation & Setup

Prerequisites

• Coding-agent host: A host with skill support, shell access, and parallel subagent spawning enabled. The coordinator launches specialists and fresh verifier agents through the host.
• Runtime access: Python and Git, plus any R, native CLI, data, API, or GPU requirements of the selected repository. The skill prepares isolated project environments and records tested versions.

Manual Installation Steps

To have your coding agent install the skill, use the Quick Start prompt. To install it yourself, follow the steps below.

1. Clone the Paper2Agent repository

   copy

   git clone https://github.com/jmiao24/Paper2Agent.git
   cd Paper2Agent
   

2. Install the entire skill folder for your host

   Choose the command for your host. Include references/, scripts/, and agents/ along with SKILL.md.

   Claude Code — personal skill location from the Claude Code skills documentation:

   copy

   mkdir -p "$HOME/.claude/skills/paper2agent"
   cp -R skills/paper2agent/. "$HOME/.claude/skills/paper2agent/"
   

   Codex — personal skill location from the official OpenAI skills documentation:

   copy

   mkdir -p "$HOME/.agents/skills/paper2agent"
   cp -R skills/paper2agent/. "$HOME/.agents/skills/paper2agent/"
   

3. Start your coding agent in your analysis workspace

   Open Claude Code or Codex in the directory where you want to work, then use the Quick Start prompt. If the skill does not appear, restart the coding agent. The skill installs the generated server's dependencies in its project environment during conversion.

Multi-agent Workflow

1. Prepare the environment and select tools concurrently.
2. Run selected upstream sources in parallel to obtain reference results.
3. Implement minimal wrappers in parallel, then launch fresh, separate agents to verify them.
4. Integrate the verified tools into an MCP server and exercise real MCP calls.
5. Install and validate the server in a fresh runtime environment.
6. Package the server and have an independent verifier check installation and tool calls from the extracted ZIP.

See the skill and orchestration instructions for the workflow and resume behavior.

🤖 How to Create a Paper Agent?

Connect the generated Paper MCP server to an AI coding agent, such as Claude Code, Codex, or the Google Gemini CLI, to use its scientific tools in conversation.

Connect a Generated Local MCP Server

Extract the delivered ZIP and follow its USAGE.md to install dependencies and configure your MCP client. The instructions include the tested interpreter, server entry point, required environment variables, and supported platforms.

To have the coding agent configure the connection, explicitly request it after conversion:

Connect the generated MCP server to my coding-agent client using its USAGE.md.

Connect a Remote MCP Server Hosted on Hugging Face

You can also use an existing server from Connectable Paper MCP Servers. Open the hosted service's instructions for its MCP endpoint, transport, and authentication requirements.

For an HTTP endpoint in Claude Code, follow the MCP connection documentation:

claude mcp add --transport http <MCP_NAME> <MCP_ENDPOINT_URL>

For example, the hosted AlphaGenome MCP server can provide tools for genomic data interpretation. Once connected, you can input a query like:

Analyze heart gene expression data with AlphaGenome MCP to identify the causal gene
for the variant chr11:116837649:T>G, associated with Hypoalphalipoproteinemia.

Verification

In Claude Code, check the server's connection status with:

claude mcp list

Or use /mcp inside Claude Code. A successful connection should appear in the server list; use a tool call to confirm the scientific workflow works with your inputs. The screenshot below illustrates the original demo connection.

📁 Output Structure

The final deliverable is <project_dir>/dist/<repo-name>-mcp.zip. It contains one server project:

<repo-name>-mcp/
├── USAGE.md                     # Installation, startup, client setup, and tool reference
├── src/
│   ├── <repo_name>_mcp.py        # MCP server entry point
│   ├── requirements.txt         # Pinned Python runtime dependencies
│   └── tools/                   # Verified tool modules and runtime helpers
└── ...                          # Required source/native runtime, licenses, and route-specific files

Exact runtime files depend on the repository. Required scientific code is included or installed from a documented, tested version. R and CLI projects include their runtime restoration or installation instructions. Any external data, models, credentials, or hardware requirements are documented in USAGE.md.

Key Output Files and Directories

Intermediate artifacts remain in the working project for inspection and resuming work:

The default ZIP contains runtime files and usage instructions. Development artifacts and examples stay in the workspace. Validation covers the delivered tools and documented conditions. See the output contract for packaging and delivery requirements.

🎬 Demos

Below, we showcase demos of AI agents created by Paper2Agent, illustrating how each agent applies the tools from its source paper to tackle scientific tasks.

🧬 AlphaGenome Agent for Genomic Data Interpretation

Example query:

Analyze heart gene expression data with AlphaGenome MCP to identify the causal gene
for the variant chr11:116837649:T>G, associated with Hypoalphalipoproteinemia.

https://github.com/user-attachments/assets/34aad25b-42b3-4feb-b418-db31066e7f7b

🗺️ TISSUE Agent for Uncertainty-Aware Spatial Transcriptomics Analysis

Example query:

Calculate the 95% prediction interval for the spatial gene expression prediction of gene Acta2 using TISSUE MCP.

This is my data:
Spatial count matrix: Spatial_count.txt
Spatial locations: Locations.txt
scRNA-seq count matrix: scRNA_count.txt

https://github.com/user-attachments/assets/2c8f6368-fa99-4e6e-b7b5-acc12f741655

🧫 Scanpy Agent for Single-Cell Data Preprocessing

Example query:

Use Scanpy MCP to preprocess and cluster the single-cell dataset pbmc_all.h5ad.

🔗 Connectable Paper MCP Servers

• AlphaGenome: https://Paper2Agent-alphagenome-mcp.hf.space
• Scanpy: https://Paper2Agent-scanpy-mcp.hf.space
• TISSUE: https://Paper2Agent-tissue-mcp.hf.space

📊 Benchmarking for Paper2Agent

For comprehensive benchmarking results and evaluation metrics of Paper2Agent, please refer to our dedicated benchmarking repository: Paper2AgentBench.

📚 Citation

@misc{miao2025paper2agent,
      title={Paper2Agent: Reimagining Research Papers As Interactive and Reliable AI Agents}, 
      author={Jiacheng Miao and Joe R. Davis and Jonathan K. Pritchard and James Zou},
      year={2025},
      eprint={2509.06917},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2509.06917}, 
}