Open Science Desktop is a local-first, model-agnostic research workbench for macOS, Windows, and Linux. It joins agents, notebooks, files, figures, reports, run records, and review into one desktop workflow, and keeps sessions, data, provenance, and outputs in local folders on the machine.
The research loop ships as agent skills. A meta-skill, ai4s-agent, chains four stages in order — explore, survey, experiment, write — and each drops an inspectable artifact into the workspace: a topic matrix and pre-survey, a 6–20 page survey PDF with citations and LaTeX source, an experiment package with runnable code and a results.json carrying provenance, and an 8–14 page paper draft with figures and tables. Separate skills render a mindmap and audit a finished paper for image, numerical, and logical problems under a four-level evidence grading.
Artifacts keep their evidence. A provenance ledger at .openscience/provenance.jsonl tracks file versions and links each figure, table, notebook, or report back to the code, inputs, environment, model output, and conversation that produced it. Command execution, dependency installs, deletion, and network access require explicit consent, and /plan lays out an execution plan before any file is touched while /goal fixes the objective and acceptance criteria.
The shell is Tauri 2 with React and TypeScript, and the agent runtime is a bundled, pinned OpenCode sidecar isolated from the user's own OpenCode configuration. osd server runs the same workspace, runtime, and web interface on a machine with no display, and the osd command drives sessions, projects, runs, files, and approvals from a terminal. A token-authenticated gateway serves the same interface to a browser on the local network or a phone; it is loopback-only until LAN access is enabled, and API keys stay on the machine.
It speaks the Agent Client Protocol in both directions: an external ACP agent can act as the runtime behind the ordinary interface, with its own model and history, or an external editor can drive Open Science over the same gateway token. The agent can also drive the user's own Chrome with its profile and logins intact, or an isolated private browser on demand.
Features
- Research skills: explore, literature survey, experiment suite, and paper writer, chained end to end or run individually
- Provenance ledger: file-version tracking that links every artifact to the run, inputs, environment, and conversation behind it
- Approval gates: command execution, dependency installs, deletion, and network access each need explicit consent
- Notebooks: real .ipynb files, Python and R kernels executed locally, a managed Jupyter environment via bundled uv, and a JupyterLab action
- Run records: append-only logs and a global SQLite run index with search, facets, pagination, output links, and reproduce prompts
- Remote compute: machines registered from ~/.ssh/config, plus SSH/Slurm, Modal, and notebook-batch run targets
- Projects and memory: named projects that import a folder in place, global and per-project memory layers, and automatic context compaction
- Headless CLI: osd server runs the workbench without a window and osd drives sessions, projects, runs, files, and approvals
- Remote access: token-authenticated gateway serving the real interface to a CLI, a LAN browser, or a phone, in read-only or full-access mode
- Editor interop: the Agent Client Protocol both ways, with MCP connectors and history replay preserved
- Science connectors: one-click MCP connectors for arXiv, PubMed, Crossref, Semantic Scholar, bioRxiv/medRxiv, ClinicalTrials.gov, MyVariant/ClinVar, Materials Project, FRED, Open-Meteo, and USGS water data
- Viewers: PDF, image, video, HTML, Markdown, code, CSV/TSV with charts, DOCX, XLSX, PPTX, molecules, 3D meshes, genome tracks, and FITS files
