Open Science Desktop (formerly Open Science) is a desktop research workbench for macOS, Windows and Linux that runs an AI agent through the whole research loop: exploration, literature survey, hypothesis, experiment code, analysis, figures and write-up. It connects agents, notebooks, files, figures, reports, runs and review in one session, and each stage leaves an inspectable artifact in the workspace rather than only a chat reply. Sessions, data, provenance, notebooks and run records stay in local folders, and nothing leaves the machine by default.
The app is a Tauri shell with a React interface that talks to a bundled, pinned OpenCode sidecar, isolated from any OpenCode setup the user already has. Models come from the OpenCode provider catalog, OAuth or API-key provider flows, and custom OpenAI-compatible endpoints. It also speaks the Agent Client Protocol (ACP) in both directions: it can run Codex, Gemini CLI, Claude Code or another ACP agent as the runtime behind its interface, and editors such as Zed, JetBrains and Neovim can drive it.
Provenance links every figure, table, report, notebook and run output to the code, inputs, environment, model output and conversation that produced it. Local, SSH and Slurm, Modal, and notebook-batch runs are recorded as reproducible run records in an append-only log with a searchable index. A headless server mode serves the same workspace, runtime and web interface on a machine with no display, and a command-line client drives sessions, projects, runs, files and approvals from a terminal or another agent. A token-authenticated gateway, off by default and loopback-only until enabled, serves the interface to a browser on the local network or a phone, and API keys never cross it.
| Skill | Role | Primary output |
|---|---|---|
| ai4s-agent | Runs the four skills below in order | The full research package |
| research-explorer | Turns a broad direction into concrete topics | Exploration notes, topic matrix, pre-survey |
| literature-survey | Writes a literature survey | 6–20 page PDF with 60+ real citations, LaTeX source and taxonomy figures |
| experiment-suite | Builds an experiment package | Design doc, runnable code, results with provenance, figures and report |
| paper-writer | Writes a research paper | 8–14 page PDF with 200+ citations, 4–8 figures and tables |
| mindmap-render | Renders a mindmap | Image generated from the topic matrix |
| integrity-auditor | Audits a paper's integrity | Image, numerical and logical findings with four-level evidence grading |
Features
- Research skills: a meta-skill chains exploration, literature survey, experiment and paper-writing skills, each producing a gradeable artifact on any model OpenCode supports
- Provenance tracking: file versions and produced artifacts link back to the run or edit that created them, and an artifact inspector shows the generating script, data and conversation
- Reproducible runs: local, SSH and Slurm, Modal and notebook-batch runs captured as run records with search, facets, logs and reproduce prompts
- Notebooks: real Jupyter notebook files in Python and R, local kernel execution and a managed Jupyter environment
- Agent Client Protocol: runs Codex, Gemini CLI, Claude Code or any ACP agent behind the interface, or lets Zed, JetBrains and Neovim drive the workbench
- Browser and computer use: drives the user's own Chrome with logins intact or an isolated browser, and operates desktop research software through the accessibility API
- Science connectors: one-click MCP connectors for arXiv, PubMed, Crossref, Semantic Scholar, bioRxiv and medRxiv, ClinicalTrials.gov, ClinVar, Materials Project, FRED, Open-Meteo and USGS water data
- Projects and memory: named projects that import existing folders in place, global and per-project memory, searchable history, and automatic context compaction
- Agent modes: a plan command lays out an execution plan before touching files, and a goal command fixes objective, constraints and acceptance criteria
- Split-pane layout: tiled sessions with drag-to-dock, independent screens and a different model per pane
- Headless server and CLI: the full workbench without a display, scriptable sessions with wait and JSON output, and approval rules that pause before commands, deletions and network access
- Viewers: PDF, image, video, HTML, Markdown, code, tables with charts, Office documents, molecules, 3D meshes, genome tracks, FITS and band-structure files
