wtf is openai4s?

OpenAI4S is an open source AI research agent built for scientific work, positioned as a much cheaper alternative to Claude Science. It can run on Doubao, a low-cost Chinese AI model service, for as little as about 9.9 yuan, roughly one and a half US dollars, per month, though it also supports ChatGPT, Claude, Gemini, and several other model providers through one shared interface. The project combines two ways of working. Structured tool calls handle predictable steps like permissions, approvals, and connecting to outside services. Separately, the agent can write and run Python or R code in a persistent session, which is where the actual scientific computation, data analysis, and simulation happen. The idea is that some tasks are better handled as clean, auditable steps, while open-ended analysis is better handled as real code that can loop, branch, and use libraries, rather than being forced through many small back and forth tool calls. It ships with a web based workbench showing live output, saved results, and a history of what the agent did, plus a notebook interface that is read-only by default unless a developer explicitly turns on manual editing. It includes 34 built in skills, many of them scientific tools for tasks like protein structure prediction and single-cell biology analysis, alongside general research workflow helpers. For heavier computation, it can dispatch jobs to remote GPU machines over SSH or through an NVIDIA integration, when one is configured. Security features include restricted environments for anything the agent runs locally, required approval steps, and operating system level sandboxing on Mac and Linux. A one click Mac app is available for Apple Silicon, bundling Python and the core scientific libraries, and it can also be installed as a Python package. The project is aimed at scientists and researchers who want an AI assistant that can genuinely run analysis code, not just describe what to do, and it is released under the MIT license.