Hi Product Hunt! I'm Gongbo, a grad student at Peking University and one of the makers of OpenAI4S. The name reads "Open AI for Scientist": we're an academic open-source project and not affiliated with OpenAI. Why we built it: most agents act by calling one JSON tool at a time. That works for booking a meeting, but research is loops, dataframes, alignments and simulations. Doing it one tool call at a time is slow, and the intermediate data ends up in the context window. How it works: OpenAI4S splits the work into two planes. JSON tool calls handle orchestration: plans, permissions, approvals, metadata and external services. The science runs as complete code cells in persistent Python and R kernels, so a 100k-row DataFrame stays in the kernel and only a short summary reaches the model. Each session keeps an append-only action ledger, versioned artifacts and a notebook you can replay. What's in the box: • Desktop packages for macOS (Apple Silicon), Linux and Windows via WSL2, with the science stack (RDKit, Scanpy, Biopython...) bundled. Or pip install openai4s. • 604 Skills, which are code recipes rather than tool schemas, including the 561-recipe bioSkills collection • Connectors for UniProt, RCSB PDB, Ensembl, ChEMBL, PubChem, arXiv and OpenAlex • Works with Doubao, GLM, Kimi, DeepSeek, MiniMax, GPT, Claude or Gemini. It is built to run on low-cost models: in China, a ¥9.9/month (about US$1.4) Volcengine Ark plan is enough. • OS sandboxing (Seatbelt on macOS, bubblewrap on Linux) and approvals for risky actions It's free and MIT-licensed. Technical report: https://arxiv.org/abs/2609.15096 We'd love feedback from people who do computational research: which workflows break, which Skills are missing, and what it would take for you to trust an agent's results enough to put them in a paper.

OpenAI4S

Open-source AI research agent that runs science as code

Open-source AI research agent that runs science as code