Stanford's Paper2Agent packages published papers as callable tools
single source· 1 articles · confidence: low · first seen 2026-09-16 21:59 UTC
What this means for you
Nothing to act on yet — no code, no release date, no harness, so the 91.2% cannot be checked or reproduced. If the method becomes usable, the test worth running is whether it re-derives a result from the methods section alone; today that claim rests on a single reported number.
Stanford researchers have released Paper2Agent, a system that reads a published research paper and rebuilds it as a validated MCP tool — a callable function an AI model can run, exposed through the Model Context Protocol. The work is published in Nature. The reported figure is 91.2% on 300 questions drawn from 74 papers; the write-up gives no evaluation date, no harness, no release date and no code link. The pitch is that reproducing a result becomes a matter of running a tool rather than reimplementing someone's code.
Key facts
What the sources say
- MarkTechPost — Short press write-up carrying the only figures available: a 91.2% score across 74 papers
Sources
The original reporting. Follow these — they did the work.