What Is Open Science? A Local-First, Model-Agnostic AI Research Workbench

Open Science is an open-source, local-first, model-agnostic, self-hosted AI research workbench for reproducible scientific discovery. Built for researchers, it runs on your own computer (macOS, Windows, Linux). Create a project, describe a task in plain language, and let the AI agent read files, run Python and R code, search the web, call scientific data connectors, and return reproducible reports, tables, and figures linked to an inspectable activity history in one workspace.

The Real Problem Researchers Face Every Day

AIPOCH Open Science addresses a core problem: research context keeps getting lost across chat windows, notebooks, file browsers, and database sites, making results hard to trace and reproduce.

Research tasks rarely finish in a single conversation. A typical project spans literature search, data cleaning, code execution, figure generation, result interpretation, and repeated iteration. Researchers often switch between 5–10 different tools. With every switch, context is lost: the model’s conclusion no longer shows the original code, the generated figure cannot be linked to its input data, and changing one prompt overwrites the previous answer.

The cost of this fragmentation is not only efficiency. More seriously, the research result becomes disconnected from the process that produced it. Collaborators—and even the original author—struggle to reconstruct “how this figure was made” weeks later. For work that needs to be reproduced, audited, or shared, this lack of traceability is a structural risk.

What Open Science Is at Its Core

Open Science is not a replacement for web chat tools. It is an AI research workbench organized around durable projects. Its positioning can be summarized in one sentence: an open-source, local-first, model-agnostic, self-hosted AI research workbench for reproducible scientific discovery.

Each word in that sentence has a concrete meaning:

• ​Model-agnostic​: it supports multiple model providers, custom gateways, and Claude / Codex subscription logins.
• ​Open-source​: the code is released under the Apache 2.0 license, so anyone can inspect, fork, or extend it.
• ​Local-first​: projects, sessions, files, artifact versions, and evidence are stored on your own computer by default.
• ​Self-hosted​: the application runs locally; you do not have to hand your research data to a cloud account.

As of August 2026, Open Science v0.10.1 supports persistent projects and sessions, multi-provider model connections, logged agent tool activity, Python/R notebooks, 18 featured scientific Skills, and 24 built-in scientific Connectors.

What Makes Up the Workbench

Open Science places a research task inside a unified project space. The key modules are projects, agents, artifacts, provenance, execution environments, skills, and connectors.

Projects and Sessions

A project is the basic container for a research effort. It holds multiple sessions, uploaded files, generated artifacts, and run history. After you close the app or restart your computer, everything is restored. A session is a continuous conversation with an agent. You can run several sessions within one project for different subtasks and pin important sessions.

More importantly, Open Science supports ​message branching​: you can edit a completed user message and resend it, and the system creates a new branch from that point while keeping the original downstream conversation intact. This lets you explore alternative hypotheses without overwriting prior work.

Agent and Tool Activity

In a session, the agent does not only return text. It invokes tools based on the task: reading files, running Python/R code, searching the web, calling scientific connectors, editing files, or generating reports. Each tool call is recorded as a card in the conversation, so you can see which files were read, which code was executed, and which external services were called. Open Science offers Ask / Auto-approve edits / Full access approval profiles, letting researchers choose the right permission level for each task’s sensitivity.

Artifacts and Provenance

Tables, figures, and reports generated by the agent are saved as Artifacts in the project file library. Each Artifact is an immutable, checksummed version with a Provenance view that shows the evidence Open Science can verify: the producing code, execution history, referenced inputs, environment inventory, producing conversation branch, and reviewer findings (if enabled). Evidence that cannot be verified is explicitly marked as unavailable rather than guessed.

Execution Environments

Open Science includes persistent Python, R, and REPL execution environments. Code written by the agent runs inside the project space, and both code and output history are preserved. You can choose an app-managed environment, an interpreter already on your machine, or a remote SSH compute host via the Remote Compute (SSH) Skill.

Skills and Connectors

A Skill is a reusable research method package. Open Science currently provides 18 featured built-in Skills, including Literature Review, AlphaFold2, Boltz, ProteinMPNN, scvi-tools, and Remote Compute (SSH). You can create personal Skills, import community Skills from GitHub, or select an enabled Skill in a session with /.

A Connector links external data or tools to the agent. Open Science currently provides 24 built-in scientific Connectors covering literature (PubMed, bioRxiv, Literature Graph), genes and ontologies, structures and functions, chemistry and drugs, clinical trials, and more. You can also add custom MCP Connectors; each connector and tool has independent permission controls.

Model Choice

Open Science supports multiple model providers, including OpenAI, Anthropic, xAI, DeepSeek, Zhipu AI (GLM), Kimi (Moonshot), MiniMax, StepFun, Xiaomi MIMO, SenseNova, Volcengine Ark, Bailian (Alibaba Cloud), and OpenRouter. You can also use a compatible custom gateway or reuse an existing Claude / Codex subscription login. Model selection is per-session, so researchers can switch based on task needs and cost preferences.

How Open Science Works

A typical workflow has five steps: create a project, upload or reference files, describe the task, review execution, and trace or branch. This turns “asking the model a question” into “completing a traceable research task.”

1. ​Create a project​: give the study a stable name; the project becomes the container for all sessions, files, and artifacts.
2. ​Upload or reference files​: upload CSV, PDF, Notebook, and other files to the project library—up to 10 GB per file—and reference them with @ in a session.
3. ​Describe the task​: tell the agent the goal, input data, constraints, expected output, and how the result should be checked.
4. ​Review execution​: the agent invokes tools and generates artifacts. You can inspect each tool call, approve sensitive actions, and review generated code.
5. ​Trace and branch​: open an Artifact’s Provenance view to inspect source evidence; if you want to try a different method, edit an earlier message to create a new branch while the original result remains intact.

Conclusion

Open Science combines open-source code, local data, model choice, and inspectable execution into a single desktop research workbench. It does not make scientific judgments for you, but it can help organize the research process—from file upload to code execution, from artifact generation to source tracing—into a reproducible, shareable record.

If you are new to Open Science, visit the Open Science product page for an overview, or go to the GitHub repository to download the latest release and complete your first setup.

This article is based on the public README and ​​*v0.10.1 Release* of ​*aipoch/open-science​; interface details may change with newer versions, so refer to the installed app as the source of truth.*

Frequently Asked Questions

Is Open Science free?

A: Yes. Open Science is open-source under the Apache 2.0 license, so the code is free to use. You only pay for the model service or infrastructure you choose.

Does my research data leave my computer?

A: Projects, sessions, files, and artifacts are stored locally by default. Data is only sent to external services when you actively choose a model provider, run a web search, or call a remote connector. API keys use the operating system’s secure credential storage when available.

Which models does Open Science support?

A: As of v0.10.1, Open Science supports OpenAI, Anthropic, xAI, DeepSeek, Zhipu AI, Kimi, MiniMax, StepFun, Xiaomi MIMO, SenseNova, Volcengine Ark, Bailian, OpenRouter, and others, plus custom gateways and Claude / Codex subscription logins.

What is the difference between a Skill and a Connector?

A: A Skill is a reusable research method package; a Connector is an interface to external data or tools such as PubMed or ChEMBL. Open Science currently provides 18 featured Skills and 24 built-in Connectors.

Can Open Science replace researcher judgment?

A: No. Open Science is a research workflow tool. All outputs require researcher review, verification, and interpretation. The Reviewer feature can provide additional checks, but it cannot replace domain-expert judgment.

Disclaimer

This article is intended for informational purposes only and does not constitute medical advice, clinical guidance, diagnostic recommendations, treatment decisions, or validated scientific conclusions. Sample data, workflow outputs, and parameter values shown are illustrative and do not represent any validated research finding or real-world result.

Open Science is a research workflow tool. It does not replace researcher judgment, and researchers remain fully responsible for evaluating the accuracy, completeness, and appropriateness of any outputs generated. All outputs require independent verification and expert interpretation before use in any research context.

References and external links in this article are provided for informational purposes. AIPOCH does not endorse and is not responsible for the content of third-party sources.