OpenAI4S: An open-source research agent with 30+ scientific skills, launched by Peking University and YuanKong AI

Claude Science, you may have heard of it.

Anthropic’s research agent, which can autonomously search for literature, write code, run analyses, and generate figures and reports, once led many to exclaim, “This is what AI for Science should look like.”

But it is proprietary.

Now, someone has replicated it—open source, under the MIT license, with zero core dependencies.

The project is called OpenAI4S (Open AI for Scientists), a research agent project officially open-sourced by Peking University and YuanKong AI Agent Joint Laboratory.

Key point: It’s not just “looks like” Claude Science—it independently reimplemented the engine, persistent core, host-RPC protocol, and security layer using an open-source approach, following the Code-as-Action philosophy.

The project was open-sourced on GitHub on July 6, 2026, under the MIT License.

Currently, OpenAI4S offers a complete research web application, dozens of built-in skills, integrated computing power access, and versioned research outputs.

MetaAI Science: https://www.chatexcel.com/homesite/openAI4S

It doesn't settle for providing an answer that merely looks correct; it aims to truly integrate AI into the research workflow, moving a problem from "posed" to "completed."

Many AI agents today are essentially doing the same thing: facing a pre-prepared "tool menu," deciding which tool to invoke, and then proceeding step by step. This approach works well for simple tasks.

But scientific research is not something that can be accomplished in just a few steps.

Query multiple databases, download data, clean data, repeatedly modify code, invoke specialized algorithms, request GPU computing power, generate charts, organize reports... Any step going wrong may require you to go back and start over.

OpenAI4S's answer is Code-as-Action, meaning "code as action."

It enables the agent to directly generate Python or R code and execute it in a continuously running kernel environment.

Looping, conditional logic, batch data processing, and plotting can be combined into a single execution; data and intermediate results do not need to be reloaded into the model at each step but are retained in the working environment.

For users, the most direct change isn't the addition of a new technical term, but rather that AI can now chain together longer, more complex tasks:

The data downloaded in the previous step can be directly passed to the next step for analysis;

Generated images, tables, and reports are automatically saved;

Researchers can also open the Notebook at any time to review the code, take over the analysis, or modify the results.

Code is action; the core is the environment.

OpenAI4S is more like a shared research workspace for humans and AI.

You provide a task that can autonomously query public scientific data sources, write and execute analysis code, leverage professional computing resources, and compile the results into images, tables, 3D structures, or Markdown reports.

The entire process doesn’t end once it’s completed. Every research output is recorded, saved, and versioned for easy future review, modification, and reuse.

For example, have it analyze a protein.

It can first retrieve authentic protein sequence and structure data, then perform feature calculation and visualization, and finally compile the analysis process, data sources, and results into a research paper.

The key point is: the data is real, and the calculations are real.

If an external service cannot be connected or the computational conditions are not met, it will clearly tell you what cannot be done. It will not secretly insert simulated data or fabricate results to force a paper to be completed.

△

Scientific research often requires repeated revisions.

In traditional workflows, researchers may need to locate the plotting script again, modify the parameters, rerun it, and export the file;

In OpenAI4S, you can directly request adjustments using natural language; the agent will regenerate the output based on existing data and code while maintaining version history.

△

A general-purpose model wants to become a research assistant, but being "smart" alone isn't enough.

It needs to know where to find professional data, how to use professional tools, and how to run complex algorithms.

OpenAI4S comes pre-loaded with over 30 research skills.

Protein structure, sequence analysis, protein design, molecular docking, single-cell analysis, literature retrieval, open data retrieval... broadly covers a range of common computational research scenarios.

Among them, 14 skills encapsulate scientific computing capabilities that require a GPU or specialized models.

Here, the Skills are not a static table of tool parameters; they are more like a set of "code recipes" that an agent can read, understand, and execute.

What data is needed, which services to call, how to write the code, and how to verify the results can all be captured in a Skill.

For future updates, such as new databases, algorithms, or experimental workflows, developers can simply continue integrating them.

In other words, this set of capabilities isn't fixed; as research needs evolve, Skills can grow accordingly.

Scientific computing also faces a very practical issue: computing power.

Tasks such as protein folding, mutation scoring, and molecular simulation cannot be accomplished by simply running a few tests on a regular laptop.

Moreover, many professional computations rely on specific models, complex software environments, and GPU resources.

OpenAI4S supports integration with your own computing environment. Local systems handle interaction and task orchestration, while GPU-intensive computations can be offloaded to your own servers. Once completed, the results return to the same research workflow.

For example, in protein structure prediction and mutation effect scoring, the real challenge is often not generating the calling code, but rather feeding the input into the correct computational environment and bringing the results back fully into the workflow.

You don’t need to cram everything onto a single computer or rewrite your entire workflow just to connect to a computing server—this is especially important for labs and research teams.

Data, models, and computing power can all remain within their own environments, while the agent connects them together.

When using inexpensive models for research, what’s the biggest concern? That they confidently make things up.

OpenAI4S has established a strict rule called the no-fabrication policy: using np.random to fake experimental results, artificially inventing "conservatism," substituting BLOSUM for ESM, or passing off simulated datasets as real data—all are prohibited.

Either provide genuine service, or report errors honestly.

So when you look back at the INS example—the sequence is genuinely pulled from UniProt, the structure is genuinely downloaded from RCSB, and if it can't connect to the internet, it honestly skips it instead of giving you a fake one.

To compute protein structures, it truly uses GPUs: host.fold SSHs into an 8×A100 machine to run single-sequence Protenix (AlphaFold3-level) inference and returns structures with pLDDT scores; host.score_mutations performs actual ESM scoring to evaluate mutations.

No GPU detected? It will throw an error immediately—no fake structures allowed.

It would rather tell you "can't calculate" than give you a fake structure that looks plausible.

For a research agent, this may be more important than "being good at everything."

The release of OpenAI4S is not just the addition of an open-source project, but also a practical step by Peking University and YuanKong AI to advance industry-academia-research collaboration.

As a milestone achievement of the joint laboratory, the development and open-sourcing of OpenAI4S would not have been possible without the continuous dedication and promotion by student Zhang Gongbo from the joint laboratory.

Zhang Gongbo's research focuses on AI for Science. As a co-first author of UniGenX, he explored a universal scientific foundation model for molecules, materials, and proteins. These research advancements in scientific intelligence have further extended to the development and deployment of OpenAI4S.

At the same time, OpenAI4S also reflects the collective efforts of team members Li Hao, Wang Yu, Liu Yuyang, Lü Liuzhenghao, Mao Yicheng, Lin Mujie, Peng Xinyao, Jiang Zhengxiang, Wang Yimi, and others, whose collaboration and dedication were essential throughout every stage—from framework design and feature development to open-source release.

YuanKong AI is an edge-side AI company focused on enabling AI to perform complex tasks locally and offline through model post-training, agent runtime, long-term memory, and edge-side inference technologies.

The company has currently established two product lines: YuanKong AI Work and YuanKong AI Science, entering the fields of intelligent office solutions and scientific discovery, respectively.

This open-source release of OpenAI4S also reflects the joint assessment of the joint laboratory and YuanKong AI regarding research agents:

Research agents should not remain limited to demos but should become actual, executable, verifiable, and scalable research infrastructure.

Therefore, this opening includes not only examples but also the core architecture, research skills, and complete applications.

Universities, research institutions, developers, and industry partners can integrate data sources, algorithms, tools, and research workflows on this foundation to jointly advance research agents into more real-world tasks.

Participate in OpenAI4S by getting it up and running.

Installation is simple—just three lines to get started:

Launch the project without needing to configure the model key first.

After entering the interface, simply fill in the details under Customize → Models.

The project also includes a complete set of real-world case studies that you can follow along with: Live API workflows, real-world analysis of human insulin (INS) data, one-click image modification, and solubility predictions for artemisinin and paclitaxel in planning mode, as well as protein engineering.

But getting the project running is just the first step.

Real scientific research spans numerous fields, including life sciences, chemistry, materials science, medicine, and earth sciences.

Each direction has its own database, specialized models, and computational processes. Such a vast capability landscape cannot be fully developed by a single team all at once.

Therefore, OpenAI4S is officially issuing a Call for Contributors to universities, research institutions, developers, and industry teams.

Currently, TuZhan Intelligence's Skills UI has provided partial support for this open-source project.

If you are familiar with a specific research area, you can develop new research Skills by integrating specialized databases, algorithmic models, and computing platforms.

If you are skilled in agents or software engineering, you can also contribute to refining the core architecture, remote computing, security mechanisms, testing, and documentation.

You can participate without writing any code.

Submitting an issue, providing a reproducible case, reporting a failed task, or letting the community know what a particular field truly needs are all valuable contributions.

If you also believe that research agents shouldn't remain stuck in demos, feel free to run it and join us in advancing it further.

Portal

https://github.com/PKU-YuanGroup/OpenAI4S

https://www.chatexcel.com/homesite/openAI4S

This article is from the WeChat public account "Quantum Bit," authored by: Focused on Frontier Technologies