PKU-YuanGroup/OpenAI4S

Open-source AI agent for scientific research. Analyze data in Python/R with Claude, GPT, Gemini, and more.

What it solves

OpenAI4S is a hybrid scientific research agent designed to automate complex research workflows, data analysis, and simulations. It addresses the need for a reliable, cost-effective research assistant that can handle both high-level task orchestration and deep, persistent computational work without requiring expensive frontier-model subscriptions.

How it works

The system operates on two distinct planes: a JSON-based control plane for managing workflows, permissions, and metadata, and a persistent Python/R kernel plane for executing scientific code. By separating these concerns, the agent uses structured tool calls for decision-making while keeping heavy computational tasks (like data processing or simulations) within persistent kernels. It supports a "Code-as-Action" paradigm where users can execute complete code cells that handle complex logic, with the agent maintaining state and recording artifacts throughout the process.

Who it’s for

It is built for scientists, researchers, and developers who need to automate data-intensive tasks, protein engineering workflows, or literature analysis. It is particularly useful for those who want to run sophisticated AI-driven research locally while maintaining strict control over data, environment provenance, and execution security.

Highlights

  • Hybrid Action Engine: Combines deterministic JSON tool orchestration with persistent Python/R kernels for scientific computation.
  • Ledger-first Runtime: Maintains an append-only record of all actions, usage, and completion data, ensuring research is reproducible and reconstructable.
  • Scientific Workbench: A web-based interface featuring live streaming, versioned artifacts, and a read-only-by-default notebook for managing research sessions.
  • Extensive Skill Library: Includes over 600 bundled "Skills" (code recipes) covering research workflows, platform operations, and specialized scientific domains like bioinformatics.
  • Hardened Execution: Supports OS-level sandboxing (via bubblewrap or Seatbelt) to ensure local code execution is secure and isolated.
  • Multi-Provider Support: Compatible with various LLM providers (OpenAI, Anthropic, Gemini, and Ark-compatible models like Doubao) via a unified interface.

Related

  • Project
    divelab/AIRS

    A collection of open-source software tools, datasets, and benchmarks for AI for Science, covering quantum, atomistic, and continuum systems.

  • 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
    synthetic-sciences/openscience

    OpenScience is an open‑source AI research workbench that lets users describe scientific tasks in plain language. The built‑in agent plans, searches literature, writes and runs code (Python/R/shell), executes experiments on local or remote compute, and produces a reproducible report, with every step logged. Available as a desktop app, web workspace, or CLI, it supports custom LLMs, a large library of domain‑specific skills, and extensible plugins.

  • Project
    java-up-up/nexus-agent

    An enterprise-grade AI agent platform that implements a complete loop from document governance and hybrid RAG retrieval to ReAct-based agentic reasoning and MCP tool integration.

  • Project
    K-Dense-AI/scientific-agent-skills

    A library of 177 scientific and research skills that provide curated documentation and workflows for AI agents to interact with scientific databases, Python packages, and research methodologies.