PKU-YuanGroup/OpenAI4S

9.9 元豆包 API 复刻 Claude Science

What it solves

OpenAI4S is a hybrid scientific research agent designed to automate complex scientific workflows. It bridges the gap between high-level orchestration and heavy-duty computation by separating deterministic tool use from persistent scientific execution in Python and R kernels, allowing researchers to perform analysis, simulation, and data retrieval without the overhead of constant LLM round-trips.

How it works

The system employs a dual-plane architecture:

  • JSON Control Plane: Uses native tool calls for orchestration, permissions, metadata management, and human approval.
  • Science Plane: Executes code in persistent Python and R kernels. This "Code-as-Action" approach allows the agent to perform complex computations, maintain state in kernel memory, and handle large datasets (e.g., 100k-row DataFrames) without flooding the LLM's context window.

It integrates with multiple LLM providers (including Doubao, OpenAI, Anthropic, and Gemini) and provides a scientific workbench for live streaming, artifact versioning, and data provenance. It also includes a large library of "Skills" (pre-defined code recipes) and connectors to major public scientific databases like UniProt and arXiv.

Who it’s for

It is built for scientists and researchers who need an AI agent capable of performing reproducible scientific analysis, protein engineering, and data-driven research across different programming environments (Python/R).

Highlights

  • Hybrid Action Engine: Combines JSON tools for workflow control and persistent kernels for scientific computation.
  • Pure-stdlib Core: The engine and web server are built using only the Python standard library to minimize dependencies.
  • llm-agnostic: Supports various providers via a single interface, including a low-cost path via Volcengine Ark.
  • Extensive Skill Library: Bundles over 600 scientific code recipes (including the bioSkills collection).
  • Scientific Workbench: Features a read-only notebook, action timelines, and specialized renderers for chemistry and genomics.
  • Hardened Execution: Implements OS-level sandboxing (Seatbelt on macOS, bubblewrap on Linux) for secure local code execution.

Related

  • Project
    synthetic-sciences/openscience

    An open-source AI workbench for scientific research that automates the research loop, including literature review, coding, and running experiments.

  • Project
    ZimoLiao/scholaraio

    An academic harness for AI agents that provides the structured evidence, library management, and repeatable workflows needed for professional research and academic writing.

  • Project
    xuzhougeng/wisp-science

    An open-source, local-first AI research workbench that integrates literature search, Python/R execution, and scientific database querying into a single project environment.

  • Project
    OpenNSWM-Lab/FAROS

    A multi-agent collaborative system for AI scientists that automates the research loop from ideation and literature retrieval to code execution, paper writing, and auditable peer review.

  • Project
    K-Dense-AI/science-superpowers

    A computational-science methodology for AI agents that enforces pre-registration and reproducibility to prevent p-hacking and ensure verifiable research findings.