What is openscience?

OpenScience is an open source AI workbench for doing scientific research. You give it a goal, and it works through the research process on its own: it reads the papers that matter, forms a hypothesis, writes and runs code, runs experiments on real compute, queries major scientific databases, and writes up what it found. It runs as a workspace in your browser and can use models from Anthropic, OpenAI, Google, and many other providers, using your own API keys, with no account required. The project describes itself as model agnostic and built for real work in machine learning, biology, physics, and chemistry. It runs the whole research loop, literature review through write-up, in one continuous session, using a default research agent plus specialists for biology, physics, and machine learning, along with critique and literature review helper agents and a read only plan mode. It ships with over 290 skills covering things like model training, evaluation, dataset work, molecular and clinical biology, chemistry, papers and LaTeX, figures, and cloud compute. The agent can also query around 30 scientific databases directly, including UniProt, PDB, and arXiv. The workspace itself includes a file tree, an editor, a terminal, session history, and inline viewing for molecules, structures, genomes, and plots, and it can be extended with plugins, custom agents, and a TypeScript software development kit. You install it with npm and then run a single command that opens the workspace in your browser, or run it once without installing using npx. The first run walks you through connecting either your own provider API keys or Synthetic Sciences own managed platform called Atlas, which bills model usage from a prepaid wallet so you do not need separate keys for each provider. Using your own keys is always free and never requires Atlas. Under the hood it runs a local server hosting the workspace interface, the agent, and its tools, with models routed per request so you can switch providers freely. The README is direct that the agent is not sandboxed: its permission system only keeps you aware of what it is doing, so anyone needing real isolation should run it inside a container or virtual machine. The project is released under the Apache License 2.0 and states it is an independent project not affiliated with Anthropic.