What is OpenScience?

OpenScience is an open-source AI workbench designed to act as a comprehensive co-scientist for research. It functions as a unified workspace that helps with literature reviews, code execution, experiment management, compute scaling, and scientific writing. By combining a research agent with a robust workbench, OpenScience allows users to describe a task in plain language. The system then plans the work, gathers evidence, runs code and experiments, and produces verifiable results. The core philosophy is to handle the real work of research, such as data cleaning and reproducing claims, while letting the researcher keep control over the core ideas and final judgment. Every step of the process is visible to ensure transparency and reproducibility.

Benefits

OpenScience offers several key advantages for researchers. It provides an integrated research workflow that handles everything from searching scientific sources to drafting methods paragraphs. The platform includes hundreds of bundled skills across biology, chemistry, physics, machine learning, and data engineering. Users can connect directly to major scientific databases like UniProt, PDB, ChEMBL, PubChem, PubMed, and arXiv. It also gives access to specialized scientific software tools for protein design and molecular modeling. The system supports multiple ways to access AI models, allowing users to use their own providers, run models locally, or choose a managed service. Performance benchmarks show it solves a high percentage of complex scientific tasks compared to other AI agents. The tool emphasizes transparency by providing a full trace of every action taken, including thoughts, searches, and code execution. Users maintain control through permission settings that range from full access to approval required for risky actions.

Use Cases

OpenScience can be used in various research scenarios. Researchers can use it for literature reviews where the agent searches and compares findings from scientific sources. It is ideal for data analysis tasks where the system inspects input data, writes and runs analysis code using Python, R, or shell kernels, and produces figures and reports. Scientists can use it to reproduce experiments by agreeing on claims and prerequisites before the agent compares measured results against predictions. The tool is also useful for scientific writing, where it can draft methods paragraphs, summarize findings, and prepare reports with evidence backing conclusions. Researchers can work on projects ranging from biology and chemistry to physics and machine learning. The platform supports different computing environments, from local laptops to large clusters with GPUs. Users can initialize a project folder, describe their work in plain language, and let the agent execute the necessary tools and parallel workers to complete the task.

Pricing

OpenScience offers flexible pricing options to suit different needs. Users can sign in with their existing OpenAI or other supported provider accounts and use their own API keys without additional fees from OpenScience. For those who prefer a managed service, there is an Ace option that operates on a pay-as-you-go basis. This service manages search and memory behind a single wallet and includes handpicked, tested models optimized for scientific agents. The Ace Wallet usage has no subscription fee. When using models served through OpenRouter, a funding fee of 5.5 percent by default is included in the displayed rate. Direct provider routes add no fee to the user's costs. The platform is licensed under the Apache License 2.0 and is an independent project not affiliated with any model provider.

Vibes

OpenScience has demonstrated strong performance in scientific benchmarks. It solved 53 out of 70 tasks on the Terminal-Bench Science benchmark, achieving a success rate of 75.7 percent. On the BiomniBench-DA benchmark, it achieved a score of 82.2 out of 50 public tasks. It also solved 10 out of 14 tasks on the Terminal-Bench 4.0 Science Subset, with a success rate of 71.4 percent. These results highlight its effectiveness in end-to-end research tasks compared to other agents like Claude Code, Codex, and Grok Build. The community sees it as a significant advancement in AI-assisted scientific research. The project acknowledges contributions from various open-source collections including skills from K-Dense Inc, Orchestra Research, and Hugging Face. The transparent and flexible approach has resonated well with the scientific community looking for tools that maintain the integrity and reproducibility of their work.

Additional Information

OpenScience is an independent project licensed under the Apache License 2.0. It is not affiliated with, endorsed by, or sponsored by any model provider. The project acknowledges contributions from various open-source collections, including skills from K-Dense Inc under the MIT license, Orchestra Research under the MIT license, and Hugging Face under the Apache-2.0 license. The tool can be installed via npm, run without installation using npx, or via a standalone installer for macOS and Linux. It is accessible through three primary interfaces: a desktop app, a browser workspace, and a terminal command-line interface. The project represents a significant advancement in AI-assisted scientific research by providing a transparent, flexible, and powerful workbench that empowers researchers to tackle complex problems efficiently.