Paper2Agent turns a research paper into an MCP server your agent runs
Published in Nature and MIT-licensed, it builds validated tools from a paper's own code — 91.2% accurate. Plus the open models trending today.
Copy markdown
A paper becomes a server, not a PDF
Paper2Agent, from Stanford and published in Nature, reads a paper's own codebase and emits a Model Context Protocol server — so your agent calls the paper's methods in plain language instead of you hand-wiring its repo.
Six stages, built on Claude Code's agent SDK
It clones the repo, spins up isolated environments, runs the tutorials, records reference outputs, then wraps each method as a validated, parameterized MCP tool. The pipeline itself runs on Claude Code's agent SDK.
91.2% accurate, and cents per query
On 100 bioRxiv computational-biology papers it hit 91.2% versus 80.3% for Claude with plain repo access. The AlphaGenome agent scored 100% on novel queries at about $0.20 and 1.6 minutes each.
MIT-licensed, with servers already hosted
The code is MIT on GitHub (jmiao24/Paper2Agent), prebuilt servers run on Hugging Face Spaces, and paper2agent.ai hosts a live one. AlphaGenome's 22 tools were generated in 45 minutes for $14.
Build this weekend: your own virtual corresponding author
Point it at a method paper you keep re-implementing — Scanpy, SAM 2, TabPFN — and get an MCP server your agent can query against your own data, instead of reverse-engineering someone's notebook.
Elsewhere: open video and Chinese models own the trending list
Today's Hugging Face trending is stacked with open weights you can pull now: MiniMax-H3 and Lightricks' LTX-2.5 for near-commercial video, plus DeepSeek-V4.1-Flash for fast vision-language reasoning.