Paper2Agent is a system that converts research papers into working AI agents by extracting their methods and code into MCP (Model Context Protocol) servers. The core workflow involves pointing the system at a paper and its associated code repository, then letting coordinator agents manage specialist agents to build tested, functional tools. You install it as a skill in coding environments like Claude Code or Codex, then invoke commands like `/paper2agent Convert <github-url> into tested MCP tools in <project-dir>`. The system handles environment isolation, dependency management, and version tracking automatically. The delivered output is a ZIP file with setup instructions, letting you connect the generated server to your coding agent and use the paper's methods directly in conversation. The practical setup requires a coding agent host with shell access and parallel subprocess spawning. You clone the Paper2Agent repo, copy the skill folder to your host's skill directory (the path differs for Claude Code versus Codex), and restart your agent. Runtime dependencies vary by paper—Python and Git are standard, but some repos need R, GPUs, or specific data files. The system documents which interpreter versions were tested and which environment variables to set for API keys, keeping credentials out of generated code. Three working examples demonstrate the approach: TISSUE for spatial transcriptomics analysis, Scanpy for single-cell preprocessing, and AlphaGenome for genomic interpretation. Each has hosted demo servers running at Hugging Face spaces so you can test without installing locally. The output agents work like regular MCP servers—you connect them to your coding agent, then ask natural language queries about the paper's domain. An AlphaGenome agent can identify causal genes from variant data, a TISSUE agent calculates prediction intervals for spatial gene expression, and a Scanpy agent preprocesses and clusters single-cell datasets. The system includes a benchmarking repository (Paper2AgentBench) for evaluating conversion quality and reliability. The whole approach trades off upfront setup complexity for the ability to turn any published method with public code into an interactive tool you can query from your IDE.