- Published on
Stanford Researchers Release Paper2Agent: Turning Research Papers Into AI Agents That Reproduce Results and Run on Ne...
- Authors
- Name
- DeepHorus
Stanford researchers have released Paper2Agent, a system that converts a scientific paper and its accompanying codebase into a Model Context Protocol (MCP) server that any compatible AI agent can operate through natural language. The work, led by Jiacheng Miao and James Zou, was published in Nature on 16 September 2026. The authors describe the resulting artifact as a "virtual corresponding author" that reproduces published results and can also run the paper's methods on new data.
Computational papers typically ship code that readers must clone, install, configure and debug before they can use it. That friction keeps useful methods effectively locked inside PDFs, limiting reproducibility and reuse. Paper2Agent targets this bottleneck by automating the entire packaging process from repository to runnable agent.
The system is MIT-licensed and installs as a skill for Claude Code or Codex. Prebuilt servers for AlphaGenome, Scanpy and TISSUE run on Hugging Face Spaces, and a hosted version is available at paper2agent.ai.
Technically, Paper2Agent runs on Claude Code's agent SDK. A central orchestrator dispatches specialized sub-agents through six steps: locating and downloading the codebase, building an isolated virtual environment, indexing usable tutorials, executing those tutorials end to end and recording reference outputs, extracting parameterized MCP tools, and verifying them. The orchestrator then assembles validated tools into one MCP server.
The validation gate is strict. A tool passes only when expected files appear and numbers match references within 3%. Figures must match by perceptual hash with Hamming distance under 20. The verifier receives up to six attempts per function, and tools that keep failing are excluded. Each server exposes MCP tools, resources and prompts, with Claude Sonnet 4 used throughout.
For AlphaGenome, Paper2Agent built 22 tools in about 45 minutes for US$14, all passing validation without human intervention. On 15 tutorial-derived queries it scored 98.7±1.3%, versus 82.7±3.4% for Claude Code with repository access and 37.3±4.0% for Biomni. On 15 novel queries it reached 100.0±0.0%, against 78.7±4.4% and 56.0±3.4% respectively. On 30 open-ended queries it scored 82.7±2.4%, compared with 56.7±2.3% and 72.2±2.2%.
Median runtime on tutorial queries fell 1.9× versus Claude plus repository access and 3.1× versus Biomni. Results span five runs graded by two human experts with 96.7% inter-rater agreement, and the gains persisted when the baseline was upgraded to Claude Opus 4.6. The agent also re-examined an LDL cholesterol variant, chr1:109274968:G>T, ranking SORT1 as the likely causal gene. The original AlphaGenome paper emphasized CELSR2 and PSRC1, though GTEx shows significant liver eQTLs for all three.
The release fits a broader push toward agentic science, where papers become executable infrastructure rather than static documents. By lowering the barrier to reproducing and extending published methods, Paper2Agent could accelerate reproducibility checks and method reuse across computational biology.
