InternAgentS – A Domestic Scientific Agent Workbench Open-Sourced by Shanghai AI Lab

1. What is InternAgentS

InternAgentS is a scientific agent workbench open-sourced by the Shanghai Artificial Intelligence Lab (Shanghai AI Lab) for AI for Science scenarios. This platform integrates previously fragmented research processes—such as paper reading, experimental analysis, code iteration, remote computing, and scientific writing—into a unified project space, forming an end-to-end scientific workflow. InternAgentS supports integration with multiple large language models, including DeepSeek, Qwen, Kimi, GLM, Intern-S, and Claude, and allows users to deploy the platform locally, thereby effectively safeguarding the privacy and security of unpublished papers and experimental data. The platform has already connected to over 3,600 scientific research tools and Skills within the SCP 2.0 ecosystem, covering six major academic disciplines: biology, chemistry, physics, materials science, earth science, and mathematics and information science. It is also compatible with the MCP protocol, offering strong scalability.

Image source: official article

Technical Positioning and Domain: InternAgentS belongs to the AI for Science domain and serves as an infrastructure platform for scientific agent systems. It is positioned as middleware connecting researchers with AI capabilities. Unlike general-purpose AI workbenches or simple code assistants, it is specifically designed to address the complex coordination needs across multiple stages, tools, and models in scientific workflows. The goal is to seamlessly integrate tasks such as literature research, experimental design, code execution, result analysis, and paper writing through agent orchestration capabilities.

Development Background: This platform was developed by the Shanghai Artificial Intelligence Lab (Shanghai AI Lab), one of the core forces in China for AI fundamental research and open-source ecosystem building. The lab has accumulated deep technical expertise through projects such as the Intern series of models and OpenMMLab. The motivation for development stems from the widespread issues in current scientific research, including fragmented tools, data silos, and reliance on a single model. The aim is to provide an open, controllable, and customizable scientific agent solution for both academia and industry.

Core Value: The core value of InternAgentS lies in solving the "breakpoints" in the scientific research process. Traditional scientific research requires frequent switching between literature management tools, code editors, computing clusters, and writing software, resulting in inefficient information flow. InternAgentS unifies the management of papers, code, data, and experimental results within a project space and leverages agents to automatically perform repetitive tasks such as literature retrieval, code modification, and experimental analysis. This frees researchers from tedious workflows and allows them to focus on core scientific questions. Additionally, the ability to deploy locally ensures data sovereignty, which is crucial for research teams dealing with unpublished results and sensitive data.

Technical Features: The technical advantages of InternAgentS lie in the combination of "multi-model flexible adaptation" and a "deep tool ecosystem." It is not tied to a single model and supports mainstream domestic and international models as well as local private models, effectively avoiding vendor lock-in. The integration with the SCP 2.0 ecosystem provides over 3,600 discipline-specific tools, far exceeding the scope of general-purpose workbenches. Furthermore, the platform supports both SCP and MCP protocols, offering standardized interfaces for community expansion and third-party tool integration, thus building an open scientific agent infrastructure.

2. Key Features

• Project Space Management: Organize papers, code, data, experimental results, scientific charts, and generated artifacts around research projects, forming a traceable and reusable research workflow. All files and intermediate results generated by the agent are managed within the same space, avoiding issues such as scattered files and version confusion, especially suitable for long-term, multi-stage research projects.

• Literature Reading and Writing: The agent assists with literature research, method comparison, report writing, and paper composition. Users can specify research directions, and the agent automatically retrieves relevant literature and extracts key information to generate structured reviews. Writing outputs are directly stored back in the project space, supporting multiple iterations and revisions.

• Experimental Analysis and Code Iteration: Supports experimental data analysis, code modification, result organization, and generation of phased conclusions. The agent can understand the logic of experimental code, automatically adjust parameters or rewrite parts of the code based on analysis needs, and generate visual result reports, significantly shortening the experimental iteration cycle.

• Remote Computing: Supports connecting to remote Linux hosts (such as lab servers or supercomputing clusters), executing computational tasks after review and authorization by the researcher. The agent is responsible for submitting jobs, monitoring the running status, and automatically retrieving and analyzing results upon completion, achieving full automation of the entire process from "code writing - submission - result analysis."

• Multi-Model Integration: Supports mainstream models such as DeepSeek, Qwen, Kimi, GLM, Intern-S, and Claude, as well as private model services deployed on local or institutional internal networks. Users can flexibly switch models based on task types (such as literature understanding, code generation, mathematical reasoning) or call multiple models simultaneously for cross-validation.

• Tool Ecosystem Integration: Integrates with over 3,600 research tools and Skills from the SCP 2.0 ecosystem, covering six major disciplines including biology, chemistry, physics, materials science, earth science, and mathematics and information science. These tools include molecular simulation software, crystal structure analysis tools, numerical simulation engines, and the agent can automatically invoke them to complete specialized computational tasks.

• Protocol Compatibility: Supports the SCP (Scientific Intelligence Context Protocol) and MCP protocols, facilitating the extension of third-party tools and custom Skills. Developers can write new tool plugins based on protocol specifications, seamlessly integrating them into the InternAgentS workflow to build highly customized research agents.

3. How to Use

1. Environment Requirements and Prerequisites: To use InternAgentS, you need a computer with a Python 3.9+ runtime environment, with a recommended memory of no less than 16GB. For scenarios requiring local execution of large models, it is recommended to equip the system with an NVIDIA GPU (with at least 8GB of VRAM) or a Huawei Ascend NPU. The operating system supports Linux (Ubuntu 20.04+ recommended), macOS, and Windows (requires WSL2 support). Additionally, Git and Docker (optional, for containerized deployment) need to be installed.

2. Access the Official Website and Clone the Repository: First, visit the official InternAgentS website (link to be updated after official release). Clone the project repository using Git: git clone (link to be updated after official release). Then, install the dependencies by running pip install -r requirements.txt.

3. Create a Project Space: After launching InternAgentS, create a new research project via the web interface or command line. Import paper PDFs, experimental code, raw data, and other research materials into the project space. The system will automatically parse the file structure and build an index, and the agent will perform subsequent tasks based on this context.

4. Configure the Model: In the project settings, select and integrate the required AI models. For cloud-based models (e.g., DeepSeek, Qwen), enter the corresponding API Key and endpoint address. For local private models, ensure that the model has been deployed using an inference framework such as vLLM or Ollama, and configure the corresponding service URL.

5. Integrate Tools: In the tool management page, configure SCP/MCP connections to integrate research tools and Skills. The platform provides pre-configured connection templates for commonly used tools, and users can also write custom tool configurations based on the SCP protocol documentation. Once the connection is successful, the agent can automatically call these tools during tasks.

6. Initiate Tasks and Review Authorization: Describe the task requirements to the agent within the project space, for example, "Find the latest literature on perovskite solar cells and generate a review." The agent will automatically plan the execution steps and generate intermediate results (such as a list of literature, code drafts), which users can review and modify. For remote computing tasks, the agent will submit the job and wait for user authorization before executing it, with the results automatically returned to the project space.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: For university laboratories and research institutions, especially for projects involving unpublished papers or sensitive data, InternAgentS is the optimal choice. Its local deployment capabilities and 3600+ discipline-specific tool ecosystem offer a level of professionalism and security that other platforms struggle to match. If the team primarily relies on Claude models and has less stringent data privacy requirements, Claude Science provides a more out-of-the-box experience. For development teams requiring highly customized agent workflows, AutoGPT and CrewAI offer more flexible frameworks, but they require the team to build their own tool ecosystem and disciplinary knowledge base, making them better suited for teams with strong AI engineering capabilities.

6. Editor's Take

The launch of InternAgentS marks a significant step forward in the field of AI for Science, transitioning from "point tools" to "platform-based infrastructure." From a technological innovation perspective, its standout feature is the integration of multi-model adaptation, deep tool ecosystems, and remote computing capabilities into a unified project space architecture, rather than merely stacking features. This design philosophy addresses a long-standing issue in scientific research—the "discontinuity" between literature management, code experimentation, computational resources, and paper writing. Information flow no longer requires manual handling, as agents now serve as "digital research assistants" connecting all these elements. Particularly noteworthy is the introduction of the SCP 2.0 protocol, which provides a standardized interface for connecting scientific research tools. This protocol is more aligned with the computational needs of specific disciplines than the general-purpose MCP protocol, reflecting a deep understanding of scientific research scenarios.

In terms of practical value, InternAgentS is especially beneficial for computationally intensive disciplines such as materials science, computational chemistry, and engineering simulation. Taking molecular simulation as an example, traditional workflows require manually writing input files, submitting computational jobs, and parsing output results. With InternAgentS, this process can be encapsulated as an agent task, automatically completing parameter optimization, job scheduling, and result analysis, reducing hours of work to minutes.

In terms of target users, InternAgentS is most suitable for research teams with a certain level of AI engineering capability, including university research groups, corporate R&D departments, and independent researchers. For scientists with no programming knowledge, the current version has a relatively high learning curve. However, with the improvement of community documentation and the optimization of graphical user interfaces, it is expected to reach a broader user base in the future.

In terms of development potential, InternAgentS is currently in a phase of rapid iteration. If it can further reduce configuration complexity, add pre-built workflow templates, and improve community collaboration mechanisms, it has the potential to become the standard infrastructure in the AI for Science domain. However, its long-term development still depends on the prosperity of the SCP ecosystem and the level of community engagement.

The deduction points are the steep learning curve of the current version and the lack of completeness in some functional documentation. However, the leading nature of its core architecture and tool ecosystem gives it a significant differentiating advantage among similar products, especially for research teams with high requirements for data security and disciplinary depth.

7. Application Scenarios

• Materials Science Analysis: Researchers can use InternAgentS for material property calculations, crystal structure analysis, and data mining. The agent automatically calls computational tools such as VASP and Quantum ESPRESSO, screening candidate materials based on experimental data to assist in the discovery and selection of new materials, significantly shortening the R&D cycle.

• Computational Chemistry Research: In the field of computational chemistry, InternAgentS can perform complex tasks such as molecular simulations, reaction pathway calculations, and quantum chemistry computations. For example, when a user describes "calculate the HOMO-LUMO gap of the caffeine molecule," the agent can automatically generate input files, submit Gaussian or ORCA computational tasks, and return result reports.

• Engineering Modeling and Simulation: Engineers can build physical models and conduct numerical simulations within InternAgentS. Typical applications include fluid dynamics analysis of Y-type microfluidic mixers and turbulence intensity simulation of wind turbines. The agent handles mesh generation, solver configuration, and post-processing, while users only need to review the final results.

• Literature Review and Report Writing: Researchers can let the agent automatically retrieve a large volume of literature, extract key methods, experimental data, and conclusions, and generate structured literature reviews and research reports. The agent supports automatic citation formatting and version iteration, making it suitable for writing sections such as paper introductions and method overviews.

• Interdisciplinary Collaborative Research: InternAgentS integrates tools and data resources from multiple disciplines, including biology, chemistry, physics, and materials science, supporting collaborative work on interdisciplinary projects. For instance, in a new drug development project, the agent can simultaneously call chemical molecular library screening tools, biological activity prediction models, and physical property calculation tools, achieving an end-to-end virtual screening workflow.

8. FAQ

Q: Is InternAgentS completely free to use?
A: InternAgentS itself is fully open-source and free, licensed under MIT for both commercial and academic use. However, calling cloud-based models (e.g., DeepSeek, Claude) during usage requires users to bear their own API call costs; if using a local private model, only hardware costs are incurred.

Q: How is the security of research data ensured?
A: InternAgentS supports fully local deployment—all data (including papers, code, and experimental results) is stored on the user’s own server or computer and never passes through any third-party cloud. For remote computing tasks, data is transmitted to designated compute nodes only after user authorization, and the transmission can be encrypted via HTTPS.

Q: Can multiple AI models be used simultaneously?
A: Yes. InternAgentS natively supports multi-model integration. Users can configure multiple models in the same project and flexibly switch between them based on task type (e.g., literature comprehension, code generation, mathematical reasoning). The agent can also automatically invoke multiple models for cross-validation of results.

Q: How can new research tools be extended?
A: InternAgentS supports both SCP and MCP protocols. Users can write tool plugins according to the protocol specifications and encapsulate them as standard tool interfaces. The platform provides detailed development documentation and sample code. Once a developer registers a plugin, the agent can automatically invoke that tool in tasks.

Q: What are the hardware requirements?
A: If only using cloud-based models, a standard laptop (8GB RAM) is sufficient to run InternAgentS’s web interface. For running large models locally, an NVIDIA GPU (8GB+ VRAM) or Huawei Ascend NPU is recommended, with an RTX 4090 or A100 suggested for a smooth experience. Remote computing tasks require a Linux host accessible via SSH.

Q: What are the main differences between InternAgentS and Claude Science?
A: The core differences lie in openness, model choice, and deployment. InternAgentS is fully open-source, supports local deployment, and can connect to multiple models; Claude Science is a closed-source cloud service limited to Claude models. Additionally, InternAgentS accesses over 3,600 discipline-specific tools via SCP 2.0, offering a far larger tool ecosystem than Claude Science.

Q: Does it support a Chinese interface and Chinese document processing?
A: Yes. InternAgentS’s web interface offers a Chinese version. The agent’s ability to understand Chinese documents and instructions depends on the underlying model. When using domestic models such as DeepSeek, Qwen, or Kimi, Chinese processing performance is excellent; with English-centric models like Claude, Chinese capability is relatively weaker.

9. Project Links

• Project Website: https://internagents.github.io/
• GitHub Repository: https://github.com/qzzqzzb/OpenClaudeScience