Stanford Paper2Agent Turns Papers Into AI Agents

Paper2Agent turns a computational research paper into a working AI agent. A Stanford team led by Jiacheng Miao and James Zou built the system, published in Nature on 16 September 2026, to solve a problem every computational biologist knows too well: papers ship code that readers must clone, install, configure, and debug before they can reproduce a single result.

The tool turns a paper and its codebase into a Model Context Protocol server. Any MCP-compatible agent, and that includes Claude Code, can then run the paper's methods through natural language, which means you don't need to touch the code yourself. The authors call the result a virtual corresponding author. That framing is ambitious. But the numbers behind Paper2Agent give it weight, and they're the reason the claim holds up.

The code 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.

Six Steps From PDF to Working Server

Paper2Agent runs on Claude Code's agent SDK. A central orchestrator dispatches specialized sub-agents through six steps. First, the system locates and downloads the codebase. An environment manager then builds an isolated virtual environment. A tutorial scanner indexes usable tutorials, and a tutorial executor runs them end to end while recording reference outputs. From there, a tool extractor turns those tutorials into parameterized MCP tools, and a test verifier validates each one. The orchestrator assembles the survivors into a single MCP server.

The validation gate is strict. It's simple. A tool passes only when expected files appear and numbers match within 3%. Figures must also match references by perceptual hash, with Hamming distance under 20, and that's a bar they don't get to argue with. The verifier gets up to six attempts per function. But tools that keep failing, after all that, are excluded from the final server.

Each finished server exposes three components. MCP tools wrap the paper's methods as executable functions, and they're the part that turns research into something you can actually run. MCP resources hold the manuscript, code links, datasets, and figures. MCP prompts encode multi-step workflows. Take the correct Scanpy preprocessing order. Every Paper2Agent application ran on Claude Sonnet 4.

AlphaGenome Agent: 22 Tools in 45 Minutes

For AlphaGenome, Paper2Agent built 22 tools in about 45 minutes for US $14. All 22 passed validation without human intervention. They're validated. The team then compared the agent against Claude Code with repository access, labeled Claude + Repo, and against Biomni, a comparison that put the agent's performance side by side with two other systems to see how it stacked up. But they don't stop there.

Across five runs graded by two human experts with 96.7% inter-rater agreement, median runtime on tutorial queries fell 1.9 times versus Claude + Repo and 3.1 times versus Biomni. The gains held even when the baseline was upgraded to Claude Opus 4.6.

Then came the interesting part. The agent re-examined an LDL cholesterol variant, chr1:109274968:G>T. It ranked SORT1 as the likely causal gene. The original AlphaGenome paper emphasized CELSR2 and PSRC1. GTEx shows strong liver eQTLs for all three genes, which the research team says illustrates how hard causal gene assignment is at such loci.

The research team say this shows how hard causal gene assignment is at such loci.

Scanpy, TISSUE, and a Scale Test

It took about 45 minutes. The Scanpy agent received seven validated tools, and the cost came to just US $13, which is a striking figure when you consider what that same work would demand from a human researcher in time, effort, and money. On four public datasets, it matched human researchers on cell counts, gene counts, and top marker genes. And it's not just Scanpy. A TISSUE agent reproduced human results on spatial transcriptomics data. That's the whole story.

Scale tests covered three corpora with no manual cleanup. Of 100 bioRxiv computational biology papers, 74 were agentified, and 593 of 599 proposed tools passed validation. On 300 questions, Paper2Agent scored 91.2%, versus 80.3% for Claude + Repo on Sonnet 4 and 86.3% on Sonnet 4.6. Cost per query came to US $0.20 and 1.6 minutes, compared with US $0.38 and 4.3 minutes. Ten non-biology papers, including TabPFN, SAM 2, and SAELens, hit 98.1% accuracy on 42 execution tasks. For 26 data-focused papers, the resource layer scored 89.0% versus 82.0% for browser use, at 34 times cheaper and 15 times faster.

Paper2Agent rejected 100% of out-of-scope queries. That's total. In a permuted benchmark, no less. And it recovered from injected dependency failures, file-path failures, typo failures, and deprecated API failures, which is the kind of quiet resilience you don't notice until something breaks and the system keeps going anyway. Knowing when not to answer may matter as much as answering.

Three Agents, One Psoriasis Hypothesis

The research team connected three agents: AlphaGenome, an MPRA-coupled scCRISPRi screen, and a CD4+ T cell Perturb-seq dataset. AlphaGenome flagged GPR137 at psoriasis locus rs887314, with an RNA-seq quantile score of 0.997. The AI co-scientist proposed 10 validation strategies, and a researcher picked signature correlation. Only GPR137 knockdown matched the CRE perturbation signature. The match appeared under stimulation: Spearman 0.613 at Stim8hr and 0.630 at Stim48hr. BAD and three other candidates showed no real correlation.

A second study paired AlphaGenome with an ADHD GWAS and nominated rs1626703 among 209 candidates. That hypothesis still needs experimental validation.

What Paper2Agent Actually Delivers

• Papers and repos become tested MCP servers with tools, resources, and prompts.
• The AlphaGenome agent took about 45 minutes, cost US $14, and scored 100% on novel queries.
• 74 of 100 bioRxiv papers were agentified, with 593 of 599 tools validated.
• Three paper agents jointly supported GPR137 as the probable psoriasis causal gene.
• The code is MIT-licensed, with prebuilt MCP servers on Hugging Face Spaces.

The pitch behind Paper2Agent is simple. Computational methods should not stay locked inside PDFs, waiting for a graduate student to fight through a broken conda environment. Whether the agent's causal gene calls survive wet-lab scrutiny is another question, and the team is upfront that one of them is still unvalidated. But as a way to make published methods runnable, Paper2Agent already has a track record worth watching.