OpenScience is an open-source AI workbench aimed at the full scientific research loop rather than at chat: you give it a goal, and it reads the literature, forms a hypothesis, writes and runs code, executes experiments on your own compute, queries scientific databases, and writes up the result in one continuous session. It installs as an npm package (npm install -g @synsci/openscience, or npx synsci) and runs a local server that hosts a browser workspace with a file tree, editor, terminal, session history, and inline rendering for molecules, structures, genomes, and plots. The agent layer is deliberately singular: one adaptive Research agent handles the task end to end, loads domain skills when useful, and delegates independent Explore or Execute work under an Off/Auto/High delegation setting instead of a fixed per-turn worker quota, with a read-only plan mode. It ships 309 bundled skills covering training stacks (DeepSpeed, PEFT, TRL), evaluation, dataset work, molecular and clinical biology, cheminformatics, papers and LaTeX, and figures, plus 42 built-in connectors to scientific databases including UniProt, PDB, Ensembl, ChEMBL, PubChem, arXiv, OpenAlex, and Semantic Scholar. The project is unusually explicit about what is and is not proven: a 54-entry capability inventory sits behind a single lifecycle tool, five experimental Python capabilities run in exact hashed local or Modal environments, ten experimental BioNeMo capabilities use bring-your-own-key NVIDIA NIM adapters, and no capability is labeled verified until a release-artifact canary passes. Models are routed per request to whichever provider you configure, so frontier APIs, eligible provider subscriptions, and local open-weight models are interchangeable, and keys stay on your machine. It is built on Bun, extends through LSP integration, MCP servers, plugins, custom agents and commands, and a TypeScript SDK, and is released under Apache-2.0 by synthetic-sciences.