Paper2Agent — Reimagining Research Papers As Interactive AI Agents

TL;DR

Paper2Agent is a Stanford-authored multi-agent coding system (MIT-licensed, Nature 2026, arXiv 2509.06917) that takes a research paper plus its code repository and auto-produces a validated MCP server exposing the paper’s methods as callable tools. An orchestrator spawns specialist sub-agents (environment setup, tool extraction, test generation, verification) that build an isolated runtime, lift tutorials/figures into typed tool signatures, and iteratively generate+run tests against the paper’s own reference outputs until the server passes. The generated server installs as a Claude Code / Codex skill or a hosted MCP endpoint, with reference deployments for AlphaGenome, Scanpy, and TISSUE on Hugging Face Spaces.

Key claims

Pipeline substrate is a coordinator that spawns parallel specialist sub-agents through the host (Claude Code / Codex) with shell access, Python/Git, and per-project isolated environments; verifier sub-agents are spun up fresh to re-test generated tools [README §Prerequisites, §Advanced Usage].
Correctness contract is the paper’s own reference outputs: the pipeline “iteratively generates and runs tests to refine and robustify the resulting MCP” against ground-truth artifacts from the source repo, with the generated MCP server plus USAGE.md as the deliverable [README §Overview, §Connect a Generated Local MCP Server].
Supports three input shapes — (a) paper + GitHub repo, (b) paper + repo + specific tutorial focus (title or notebook URL), (c) repo-only with API key read from a named environment variable (credentials stay outside generated code / notebooks / ZIP) [README §Advanced Usage].
Reference deployments are hosted MCP servers on Hugging Face Spaces for AlphaGenome (genomic variant interpretation), Scanpy (single-cell preprocessing + clustering), and TISSUE (uncertainty-calibrated spatial transcriptomics), each exposing tutorial-grade analyses as MCP tool calls that an MCP-compatible agent (Claude Code / Codex / Gemini CLI) can invoke [README §Connectable Paper MCP Servers, §Demos].
Companion Nature methods section reports a large-scale ablation: removing all executable tutorials from POP-TOOLS while keeping README + source code still produces a functional MCP server whose tools reproduce human-executed POP-TOOLS CLI outputs across five analysis tasks [Nature §Large-scale evaluation of Paper2Agent].
Code + benchmarking artifacts are MIT-licensed; a hosted conversion service runs at paper2agent.ai [README §Citation].

Method

The orchestrator receives a paper (URL or local file), optionally a code repository, and an output directory, then launches specialist sub-agents in parallel through the host’s subagent-spawn mechanism (Claude Code’s Task tool, Codex’s equivalent). Each specialist owns one slice of the conversion — environment preparation, tutorial/figure/method extraction, tool-signature synthesis, test authoring, test execution, iterative repair — and they coordinate through the project directory’s filesystem rather than direct messaging. Verifier agents are spun up fresh (new context, clean prompt) specifically to re-run and audit generated tests, isolating the test-generation agent from grading its own output.

The output is a self-contained project with an installed server, pinned dependencies, a USAGE.md describing the tested interpreter / server entry point / required env vars / supported platforms, plus (optionally) a ZIP for distribution. The generated server speaks MCP over HTTP or stdio, so any MCP-compatible agent client (Claude Code, Codex, Gemini CLI) can connect it alongside other tools. Credentials never land in generated code, notebooks, reports, or the ZIP — the host’s secret mechanism or process environment stays authoritative.

Results

The README reports three reference deployments and one benchmark pointer:

AlphaGenome Agent — hosted at Paper2Agent-alphagenome-mcp.hf.space; answers queries like identifying the causal gene for a specific variant (chr11:116837649:T>G, Hypoalphalipoproteinemia) by composing AlphaGenome tool calls [README §Demos].
TISSUE Agent — hosted at Paper2Agent-tissue-mcp.hf.space; computes 95% prediction intervals for spatial gene expression (e.g. gene Acta2) from Spatial_count / Locations / scRNA_count inputs [README §Demos].
Scanpy Agent — hosted at Paper2Agent-scanpy-mcp.hf.space; runs preprocessing + clustering workflows on .h5ad single-cell datasets [README §Demos].
No-tutorials ablation — removing all executable tutorials from POP-TOOLS still yields a functional MCP that reproduces human CLI outputs on five analysis tasks [Nature §Methods].

Headline external numbers (overall paper-conversion success rate, time-to-first-tool, cross-model comparison, MCP-tool accuracy vs human-executed baseline) live in the Nature paper (10.1038/s41586-026-11044-y) and the arXiv preprint (2509.06917) rather than the README — if Luma wants to pin accuracy claims, those are the companion artifacts to file separately.

Why it’s interesting

Paper2Agent generalizes the “paper as artifact” assumption that most of this wiki runs on: it treats every downstream reader as an agent, and takes the “turn the paper into a tool” step that researchers currently do by hand (clone repo, fight environment, port the tutorial into a notebook, wire it into whatever) and makes it the output of an automated pipeline. For the Tool-Use Agents design space it adds a paper-grounded tool-provenance axis orthogonal to the sequential-depth (MiroThinker), parallel-orchestration (K2.5 PARL), and context-space (Training-Free GRPO) poles already tracked — the question is no longer “how do I train an agent to use a growing tool library” but “where does the tool library come from in the first place”. Natural sibling to Kosmos: An AI Scientist for Autonomous Discovery (Kosmos), which produces human-audit-ready reports from the reader side of the paper↔agent interface, where Paper2Agent builds the writer side; complementary to ENPIRE: Agentic Robot Policy Self-Improvement in the Real World / ASPIRE — self-evolving sensorimotor skill library that compounds across robots and the sim-to-real gap (Jim Fan / NVIDIA GEAR) (ASPIRE) which bootstrap robot skill libraries by having agents write their own primitives — Paper2Agent does the equivalent for scientific-method primitives mined from already-published code. Also a clean instance of the “multi-agent SWE pipeline shipping as a Claude Code / Codex skill” deployment shape that DimOS — The Agentive Operating System for Physical Space (v0.0.13) DimOS, InstinctFlash — High-Performance Serving Runtime for Robotics Models InstinctFlash, and Previewing the Model Hardware Standard MHS are converging on from different directions.