openscience

OpenScience is an open-source AI workbench for scientific research, distributed under Apache 2.0. You describe a task in plain language; the agent plans, gathers evidence, runs code and experiments, and returns results alongside a trace of what it thought, searched, ran and wrote. Interfaces and install It runs as a self-updating desktop app for macOS, Windows and Linux, as a browser workspace, or as a terminal command. The CLI installs via npm (`npm install -g @synsci/openscience`), can be run without installing through `npx synsci`, or via a standalone installer script on macOS and Linux. A project folder is opened with `openscience ~/research/my-project`; single turns can be issued with `openscience run` and continued with `--continue`. A `/plan` mode lets you agree on method before execution. Capabilities - Tools: shell, Python and R kernels, notebooks, a file system with explicit read/write grants, and remote compute (Modal is named for GPUs and long jobs, with each dispatch approved before running). - Scientific reach: hundreds of bundled skills spanning biology, chemistry, physics, ML and data engineering, plus connectors to databases such as ChEMBL, UniProt, PubMed and arXiv. - Delegation: a lead agent can hand bounded work to parallel workers while retaining synthesis and final say. - Extensibility: MCP servers, custom agents and commands, plugins, and a TypeScript SDK generated from the server's OpenAPI contract. - NVIDIA BioNeMo: ten bring-your-own-key adapters (Boltz-2, DiffDock, Evo 2, GenMol, MolMIM, MSA Search, OpenFold2, OpenFold3, ProteinMPNN, RFdiffusion) with strict request schemas, per-dispatch approval and hashed artifacts; marked experimental and requiring your own NVIDIA API key. Model access Bring your own provider API key or sign-in, run a local model through Ollama, LM Studio or another compatible endpoint, or use the managed pay-as-you-go option (Ace), billed at provider cost plus a 5.5% fee per request. An account is optional for your own keys and local models. Permissions and safety Permission levels range from always asking to full access; network commands ask once per destination host, and files outside the project require an explicit grant. Caches and throwaway output stay in per-session scratch space. Publishing actions such as `git push`, releases and uploads run from the lead session using the machine's own GitHub and Hugging Face logins, with no tokens requested in chat. Repository layout and development The codebase separates `backend/cli` (CLI and local server: sessions, tools, providers, skills), `frontend/workspace` (SolidJS browser workspace embedded into the CLI at build time), `frontend/ui`, `frontend/desktop` (Electron shell with signed self-updater), `frontend/docs`, `tooling/sdk`, `tooling/plugin`, and `docs/notes`. Development uses Bun: `bun run setup`, `bun dev`, `bun run check` (format, typecheck, unit suites). ARCHITECTURE.md, CONTRIBUTING.md and AGENTS.md document structure, contribution loops and conventions. Releases and community Stable releases are cut from `main` after a full rehearsal at the same commit, including packaged end-to-end tests, OS smokes and scientific capability canaries on every native platform. GitHub Releases carries desktop installers, CLI archives and checksums; the desktop app self-updates and the CLI upgrades via `openscience upgrade` or npm. Issues, security reports and conduct are handled through the corresponding repository documents. Acknowledgements OpenScience states it is inspired by OpenCode by Anomaly. Most bundled skills come from open collections, including Scientific Agent Skills and Claude Scientific Writer by K-Dense Inc. (MIT), AI Research Skills by Orchestra Research (MIT), Hugging Face skills (Apache-2.0), Anthropic's document skills, the NVIDIA BioNeMo Agent Toolkit (CC-BY-4.0 / Apache-2.0), MarkItDown by Microsoft (MIT), Claude-Science-System-Prompts, pacsomatic and Iconoir icons. Attribution files list derived skills with upstream paths and licenses. The project states it is independent and not affiliated with or endorsed by any model provider.