# Paper2Agent: Converting Research Papers into MCP Tools for AI Coding Agents > Paper2Agent is a multi-agent system that reads a scientific paper and its associated code repository, then produces a tested MCP server and accompanying skill that let AI coding agents call the paper's methods as tools. It is designed to be invoked from inside Claude Code or Codex through a single prompt. **jmiao24/Paper2Agent** — Paper2Agent is a multi-agent AI system that automatically transforms research papers into interactive AI agents. - Repository: https://github.com/jmiao24/Paper2Agent - Stars: 3,625 · Forks: 539 - Language: Python - License: MIT - Published: 2026-09-16 · Updated: 2026-09-16 · Language: en - Canonical page: https://hysenlabs.com/projects/jmiao24-paper2agent ## What Paper2Agent Does and Who It Is For Scientific research papers are often accompanied by code repositories, but calling that code from an AI coding agent requires writing an MCP server that exposes the methods as callable tools. Paper2Agent automates that conversion step. The primary users are researchers who want to ask their coding agent questions that require running paper-specific analysis (such as querying a genomics model with custom data) and engineers building AI pipelines that need to invoke scientific methods programmatically. The README links to a hosted example where an AlphaGenome MCP server can analyze heart gene expression data to identify a causal gene for a genomic variant, called by pasting a plain-text query into the agent. The system coordinates parallel specialist agents: one agent reads the paper and code, others generate and test the MCP tools, and a verifier agent confirms the tools work. The output is a ZIP archive containing the server, a USAGE.md with connection instructions, tested interpreter details, and required environment variables. ## Installing the Paper2Agent Skill The recommended installation path is to ask the coding agent to install the skill by pasting a single prompt: ```text Read https://github.com/jmiao24/Paper2Agent and install the paper2agent skill from skills/paper2agent for this coding agent. Use the paper2agent skill to agentify this paper and its associated files, alongside its code repository if available. Follow the skill instructions for the workflow, verification, and final delivery. Paper and associated files: Code repository (if available): Output directory: ``` For manual installation on Claude Code, the repository is cloned first and the skill folder is copied to the personal skills directory: ```bash git clone https://github.com/jmiao24/Paper2Agent.git cd Paper2Agent ``` ```bash mkdir -p "$HOME/.claude/skills/paper2agent" cp -R skills/paper2agent/. "$HOME/.claude/skills/paper2agent/" ``` For Codex the destination path differs: $HOME/.agents/skills/paper2agent. The README states that if the skill does not appear after installation, restarting the coding agent resolves it. ## Targeted Conversions and API Key Handling When a repository has multiple tutorials or scientific tasks, the default conversion may include more tools than needed. The README shows how to focus the conversion on a specific tutorial or task: ```text Use the paper2agent skill to convert into MCP tools in . Focus on . ``` For papers whose code requires an API key, the key is made available through the host's secret mechanism or as a process environment variable, and the agent is told the variable name: ```text Use the paper2agent skill to convert into MCP tools in . Read the required API key from the environment variable . ``` The README states explicitly that credentials are kept outside generated code, notebooks, reports, and the delivered ZIP. This is a meaningful constraint: the generated server references the environment variable by name rather than embedding the key value. ## Connecting the Generated MCP Server to a Coding Agent After conversion, the delivered ZIP contains a USAGE.md with instructions for installing dependencies and configuring the MCP client. The README recommends asking the coding agent to configure the connection automatically: ```text Connect the generated MCP server to my coding-agent client using its USAGE.md. ``` For remote servers hosted on Hugging Face, the connection uses the Claude Code MCP add command: ```bash claude mcp add --transport http ``` To verify the server is connected in Claude Code: ```bash claude mcp list ``` Alternatively, /mcp inside Claude Code shows the connection status. The README documents three Connectable Paper MCP Servers in its table, including hosted versions for TISSUE, Scanpy, and AlphaGenome, each with its own Hugging Face endpoint. ## Prerequisites and Failure Modes The README lists the requirements under Installation. The coding agent host must support skills, shell access, and parallel subagent spawning. The runtime environment must have Python and Git, plus any additional requirements of the target repository such as R, native CLI tools, GPU drivers, or API credentials. The most common failure mode is a mismatch between the environment where the skill tested the server and the environment where the user tries to connect it. The skill records tested interpreter versions and dependencies in the generated USAGE.md, but if the local Python version or a native library differs significantly, the server may fail to start. The README does not document a rollback or re-generation workflow for this case. Paper repositories that depend on large proprietary datasets or require GPU access for even a small test run can cause the verifier agent to time out or produce a server that connects but cannot complete tool calls in a standard developer environment. ## Comparison with Manual MCP Server Writing Writing an MCP server by hand for a scientific Python library involves reading the API, deciding which functions to expose as tools, writing the JSON schema for each tool's inputs, handling errors, writing tests, and writing the USAGE.md. For a library with ten to twenty callable methods, that is several hours of work. Paper2Agent trades developer time for agent compute time: it runs a multi-agent workflow that takes longer than a single prompt but produces a tested server without the developer reading the full paper or library documentation manually. The trade-off works best when the paper has a well-structured code repository and clear tutorial examples, because the specialist agents rely on those as input. Papers with only a PDF and no associated code are described as a supported input type in the README, but the quality of the output depends on how much implementation detail the paper itself provides. ## Repository Layout and License The top-level repository entries include assets/, logo/, scripts/, and skills/. The skill instructions that drive the conversion workflow live in skills/paper2agent/SKILL.md. The README points to that file for the full workflow, supported inputs, and deliverables specification. The repository is licensed under MIT. The last push was on 2026-09-17. There are no GitHub releases; the skill is distributed by cloning the repository and copying the skills/ folder. ## Conclusion Paper2Agent suits researchers and engineers who regularly work with scientific codebases and want to call those methods from inside an AI coding agent without writing wrapper code by hand. It requires a coding agent host with skill support, shell access, and parallel subagent spawning enabled, plus Python and Git in the runtime environment. Before running it on a paper that requires a GPU, proprietary dataset, or non-standard runtime, verify those dependencies are available in the working environment; the skill records tested versions in the delivered USAGE.md, but environment mismatches can still prevent the server from connecting. The last push was on 2026-09-17. ## FAQ ### What coding agents are compatible with Paper2Agent? The README describes compatibility with Claude Code, Codex, and the Google Gemini CLI. The host must support skill installation, shell access, and parallel subagent spawning. The README's manual installation instructions give separate paths for Claude Code ($HOME/.claude/skills/paper2agent) and Codex ($HOME/.agents/skills/paper2agent). ### Does Paper2Agent require access to the paper's code repository? The README lists a code repository as an optional input. When one is available, it is passed alongside the paper URL or local files. The README notes that output quality depends on the repository's structure and available tutorials, but a PDF alone is listed as a supported starting point. ### What does Paper2Agent produce after converting a research paper? Paper2Agent delivers a ZIP archive containing an MCP server, a USAGE.md with tested interpreter details, required environment variables, server entry point, and supported platforms, plus instructions for connecting the server to a coding agent client. ## Sources - [Issues](https://github.com/jmiao24/Paper2Agent/issues) - [jmiao24/Paper2Agent on GitHub](https://github.com/jmiao24/Paper2Agent) - [License: MIT](https://github.com/jmiao24/Paper2Agent/blob/main/LICENSE) - [README](https://github.com/jmiao24/Paper2Agent/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/jmiao24-paper2agent