Stanford researchers release Paper2Agent turning papers into AI agents that reproduce results

and run on new data Computational papers ship with code that readers must clone, install, configure and debug — a cost that locks useful methods inside PDF. A Stanford team led by Jiacheng Miao and James Zou proposes a fix. Paper2Agent, published in Nature on September 16, 2026, converts a paper and its codebase into a Model Context Protocol server. Any MCP-compatible agent, such as Claude Code, can then execute the paper's methods through natural language. The authors describe the result as a "virtual corresponding author."

The pipeline runs on the Claude Code agent SDK. A central orchestrator dispatches specialized sub-agents through six stages: locate and download the codebase, build an isolated virtual environment, index usable tutorials, run them end to end and document reference outputs, transform the tutorials into parameterized MCP tools with validation, and finally assemble the validated tools into a single MCP server. The validation gate is strict: a tool passes only when expected files appear and numerical outputs match within 3%; figures must match the reference by perceptual hash with a Hamming distance below 20. The validator receives up to six attempts per function; tools that fail repeatedly are excluded from the final server. Each server exposes three components: MCP tools that wrap the paper's methods as callable functions, MCP resources that hold the manuscript, code links, datasets and figures, and MCP prompts that encode multi-step workflows — for example, the correct preprocessing order in Scanpy. The team used Claude Sonnet 4 for every Paper2Agent implementation.

The AlphaGenome agent produced 22 tools in roughly 45 minutes at a cost of $14. All 22 passed validation without human intervention. Benchmarks compared it against Claude Code with repository access (Claude + Repo) and against Biomni. Results spanned five runs ranked by two human experts with 96.7% inter-rater agreement. On tutorial queries, median runtime fell 1.9× versus Claude + Repo and 3.1× versus Biomni; the gap persisted when the base model was upgraded to Claude Opus 4.6. The agent also re-examined the LDL cholesterol variant chr1:109274968:G>T and ranked SORT1 as a plausible causal gene, whereas the original AlphaGenome paper highlighted CELSR2 and PSRC1. GTEx shows significant liver eQTLs for all three genes, illustrating how difficult causal gene assignment remains at such loci. The Scanpy agent yielded seven validated tools in roughly 45 minutes at $13 and matched human researcher results on cell counts, gene counts and top marker genes across four public datasets. The TISSUE agent reproduced human results on spatial transcriptomics data.

Scaling tests covered three corpora without manual curation. One hundred bioRxiv computational biology papers produced 74 agentified papers and 593 of 599 proposed tools validated. On 300 questions, Paper2Agent scored 91.2% versus 80.3% for Sonnet 4 and 86.3% for Sonnet 4.6 on Claude + Repo. Cost per query was $0.20 and 1.6 minutes versus $0.38 and 4.3 minutes. Ten non-biology papers including TabPFN, SAM 2 and SAELens delivered 98.1% accuracy on 42 execution tasks. 26 data-centric papers showed a resource-layer score of 89% versus 82% for browser use, 34× cheaper and 15× faster. Paper2Agent also rejected 100% of out-of-scope queries in a formatted benchmark and recovered from injected dependency failures, file-path errors, typos and deprecated API. The team chained three agents — AlphaGenome, a paired MPRA scCRISPRi screen, and a CD4+ T-cell perturbation screen — and demonstrated inter-agent collaboration. The code is released under the MIT license and available as a skill for Claude Code or Codex; ready-made AlphaGenome, Scanpy and TISSUE servers run on Hugging Face Spaces, and a hosted version is accessible at paper2agent.ai.