Stanford researchers published a framework in Nature on September 16 that converts computational research papers into executable AI agents. Not a RAG system that summarizes papers — actual agents that run the paper’s code, reproduce its results, and answer novel questions about its methods. The framework, Paper2Agent, works by wrapping a paper’s codebase as a Model Context Protocol (MCP) server and exposing it to LLM-based coding agents like Claude Code. In testing on 100 computational biology papers, 74 were successfully converted. On a 300-question benchmark, Paper2Agent using Claude Sonnet 4 hit 91.2% accuracy — nearly 11 percentage points ahead of giving Claude direct repository access.
Papers as Callable Tools #
The technical core is straightforward: Paper2Agent dispatches parallel specialist agents to analyze a paper’s repository, extract its core methods, wrap them as MCP tools, validate them, and deploy a remote MCP server. From that point, any MCP-compatible client — Claude Code, Codex, or any other tool that speaks the protocol — can call the paper’s methods like functions.
What gets exposed is more than raw code. Each MCP server includes structured prompts encoding how to chain tools together, covering both the paper’s original workflows and open-ended user queries. The paper’s supplementary materials, datasets, and the manuscript itself become queryable context. The result is what the authors call a “virtual corresponding author”: ask it to apply the paper’s method to your data, and it runs the code.
Installation is intentionally minimal. Point a coding agent at a paper’s arXiv ID or GitHub repo, ask it to install the paper2agent skill, and the pipeline runs from there. Dependencies are: a coding agent with shell access, Python, and Git — plus whatever the paper itself needs (R, GPU, specific data files).
The Numbers #
The team evaluated 100 bioRxiv computational biology papers, sampled without filtering for documentation quality. The results:
- 74 papers successfully converted into validated agents
- 593 of 599 proposed tools passed automated validation
- 91.2% ± 1.6% accuracy on 300 tutorial-derived benchmark questions (Sonnet 4)
- 80.3% ± 2.3% for Claude Code with direct repository access — same model, no Paper2Agent
The AlphaGenome case study stands out: the Paper2Agent version scored 100% on novel queries and ran roughly 1.9x faster than a direct-repo baseline. On a separate generalization test — 42 execution tasks across AI, statistics, econometrics, game theory, and astrophysics — accuracy was 98.1% ± 0.8%.
What the 26% Failure Rate Actually Means #
The 26 papers that failed are worth more attention than they usually get. They didn’t fail because the framework is brittle — they failed because those papers never published usable artifacts: missing executable code, unavailable datasets, broken environments, scripts tied to a specific machine. Paper2Agent can repair broken dependencies, outdated API calls, and file path errors automatically. It cannot reconstruct code that was never published.
That 26% failure rate is, in a roundabout way, useful data. It’s a documented sample of how bad research artifact hygiene still is in 2026. The 74% success rate on unfiltered papers is higher than most researchers would have predicted.
Early Ecosystem Signals #
The framework was published nine days ago, and community packaging repos have already appeared. The awesome-paper2agent repository collects pre-built paper MCP packages — each recording the source paper, upstream repo, exact commit, license, and checksum. OmicOS, a bioinformatics platform, has already integrated Paper2Agent so that installed papers become callable tools in the conversation interface.
This points toward something like a paper package manager: install a method the way you install a library. The MCP standardization is what makes it tractable — because MCP already connects LLM agents to external tools, adding a paper agent requires no new client-side infrastructure.
The Broader Shift #
Research papers turning into callable MCP servers is not an incremental improvement. It’s a format shift — from document to runtime. The reproducibility crisis has been a known problem for a decade. Paper2Agent doesn’t solve it (26% of papers still fail), but it changes the incentive structure: papers with clean, published artifacts become dramatically more useful than papers without. That’s a forcing function the academic community has been missing.
For developers, the near-term use case is concrete: any ML, bioinformatics, or statistical method published with a GitHub repo is potentially one install away from being callable. The arXiv paper and GitHub repo are public. The Nature publication has the full benchmark details.