On 16 September 2026, a Stanford-led team published Paper2Agent, an automated framework that converts a research paper—and its associated code, datasets, and workflows—into an interactive AI agent that behaves like a “virtual corresponding author.” Instead of leaving readers to clone repos and wire APIs by hand, the system builds a Model Context Protocol (MCP) server that chat agents can call in natural language.

The open-access paper lists Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard, and James Zou among the authors. Demonstrations center on computational biology (AlphaGenome, Scanpy, TISSUE) and include larger batch tests beyond that domain.

Confirmed

  • What it builds: Paper2Agent analyses a paper and codebase with multiple specialized agents, constructs an MCP server with tools (executable methods), resources (manuscript, code, supplements), and prompts (multi-step workflows), then auto-generates tests and keeps only tools that pass validation against reference outputs.
  • How users talk to it: paper MCPs can be connected to a chat agent (the authors use Claude Code / Claude Sonnet 4 in their case studies). Any compatible LLM or agent that speaks MCP can invoke the tools without custom glue code. Servers can be hosted remotely (for example on Hugging Face Spaces).
  • AlphaGenome case: Paper2Agent produced 22 MCP tools, all of which passed automated validation, in about 45 minutes at roughly US$14 on a personal laptop with no human intervention. On graded benchmarks (5 runs), the agent scored 98.7 ± 1.3% on 15 tutorial-derived queries and 100.0 ± 0.0% on 15 novel queries, versus Claude with direct repository access (82.7% / 78.7%) and Biomni (37.3% / 56.0%). On 30 open-ended researcher-style queries it reached 82.7 ± 2.4% (Claude+Repo 56.7%; Biomni 72.2%), with lower median runtime than both baselines.
  • Scanpy case: seven preprocessing/clustering tools, all validated, in about 45 minutes for about US$13. MCP prompts encode the standard pipeline order so users mainly supply a data path.
  • Scale test: of 100 computational-biology papers processed end-to-end without manual cleanup, 74 were successfully agentified (599 proposed tools; 593 passed validation). On 300 tutorial-derived questions, Paper2Agent with Sonnet 4 hit 91.2 ± 1.6% accuracy versus Claude Code + repo with Sonnet 4 at 80.3 ± 2.3%. Query cost/latency reported at about US$0.20 and 1.6 min per query versus US$0.38 and 4.3 min for Sonnet 4 on paper+repo. Across 42 execution tasks from 10 non-biology computational papers: 98.1 ± 0.8%. For 26 data/discovery papers where executable tools could not be built, a resource-only layer scored 89.0 ± 3.1% on synthesis questions.
  • Multi-agent discovery demo: agents from AlphaGenome plus two perturbation papers were combined to prioritize GPR137 at a psoriasis-associated locus, with human selection among agent-proposed validation strategies.
  • Failure modes (authors): missing executable code, missing data/model artefacts, environment/dependency failures, and non-generalizable scripts blocked agentification for many of the remaining papers.
  • Public artefacts: the team released open-source code on GitHub (jmiao24/Paper2Agent) and a live demo that chats about the paper and can reproduce selected workflows (paper2agent.ai/live). The repo also points to hosted AlphaGenome, Scanpy, and TISSUE MCP endpoints on Hugging Face Spaces.

Unknown

  • Product maturity: code and a live demo are public, but that is not a turnkey SLA for converting every journal PDF. How widely journals will require “agent availability” artefacts remains policy speculation in the Discussion.
  • Independence of grading: AlphaGenome accuracy uses human-expert rubrics (authors report 96.7% inter-rater agreement). Broader field replication outside the Stanford evaluation stack is not in the paper.
  • Causal-gene claims: the psoriasis GPR137 result is a demonstration that agents can recombine papers; the authors themselves flag activation-dependent correlations and the hard multi-eQTL problem at GWAS loci. Treat it as a workflow showcase, not a clinical finding.
  • Security, IP, and attribution: exposing runnable paper tools raises abuse and credit issues the authors note but do not fully solve in the main text.
  • Open-ended science: the paper states hypothesis generation and mechanistic interpretation remain human-in-the-loop; agreement with a single reference answer measures faithful execution more than scientific validity when multiple answers are defensible.

Our take

Paper2Agent’s real bet is packaging: MCP turns “here is our GitHub” into something a frontier chat agent can call without re-deriving the API each time. The AlphaGenome and batch numbers are strong author-reported evidence that tool-locking beats dumping a repo into a general agent—especially when the alternative is Biomni-style breadth without paper-native tools. The harder Brocker question is upstream: if a quarter of bio papers still cannot be agentified cleanly, agentification becomes another reproducibility stress test, not a free upgrade for the literature.

Sources