OpenScience

An open-source AI scientific research workbench, which the official says can read papers, write code and run experiments. It has built-in free models and also supports connecting to models from any manufacturer, and links to more than 40 scientific databases such as UniProt and PDB.

AI Research Tools#Research Analysis#Data Insights#AI Search

Updated: 2026-10-26

What is OpenScience?

OpenScience is an open-source AI research workbench, which, according to its official statement, is supported by Y Combinator. Positioned as an AI co-scientist, it is a research collaboration agent that can read papers, write code, and run experiments with you, integrating literature, code, experiments, computing power, and results into a single workspace. It does not lock users to a specific model vendor: it comes with free built-in models, and also allows access to any vendor's models such as Claude, GPT, Gemini, as well as your own API keys. Compared with ordinary conversational AI, its difference lies in that it can directly perform hands-on operations in your workspace such as retrieval, script running, and image generation, instead of only providing text suggestions.

Who is it for?

Graduate students and postdocs in experimental disciplines who need to read papers in batches and conduct data cross-verification
Researchers in bioinformatics, chemistry, pharmacy and other fields that rely on data retrieval and comparison from public databases
University research teams that want to streamline literature searching, script writing, analysis running, and figure generation into a single workflow
Authors who need to quickly draft methodology sections with proper citations
A computing platform administrator who is used to command-line and script workflows and wants to equip the lab with a self-hostable assistant

Key features

Literature Reading & Verification

The official states that the Agent can read papers and search scientific databases. In the interface examples, it first locates and compares datasets, then cross-checks literature entries, finally provides conclusions with citations, and proactively suggests independent verification methods for the next step.

Code and Experiment Execution

The workspace has built-in file, terminal and computing power panels. The Agent can directly run scripts to process structural files and generate charts. The official claims that it covers the entire process from hypothesis to chart generation.

Access to Scientific Databases

The official states that it has built-in more than 40 scientific databases such as UniProt, PDB, ChEMBL, PubChem, and arXiv, which serve as ready-made data sources for the Agent to verify facts and retrieve data without having to configure interfaces one by one.

Free Access to Multiple Models

Model-agnostic architecture, free models are ready to use out of the box. It also supports connecting to models from any vendor and using your own API keys, allowing you to switch models according to task difficulty.

Integrated Workspace

Panels for search, workspace, files, sessions, computing power and Autoresearch are centralized on the same interface. Results can be copied and forked, facilitating re-running and iteration.

How to use OpenScience

The official website provides a one-click installation script and an npm installation method.
First, use the official sample tasks, such as comparing the measured values of the database for stability prediction, to familiarize yourself with the Agent's retrieval and plotting methods, and then switch to your own research topics
When describing a task, specify the comparison datasets, samples, and expected outputs.
This will make the Agent's analysis more aligned with your needs and reduce the need for back-and-forth clarifications.
The values and citations provided by the Agent must be verified against the original databases or papers before being written into the official manuscript.
machine-generated results cannot be directly treated as research conclusions.
It is recommended to include results and scripts in your own version control.
if the project is open-source and involves sensitive data, you can deploy it yourself instead of using a public environment.
Use computing power and session resources on demand.
break long tasks into smaller tasks with clear objectives to facilitate checking intermediate results.