Claude Science, But It Lives on Your Hard Drive

What it does

Open Science Desktop wraps the entire research lifecycle into a single Tauri-based desktop app for macOS, Windows, and Linux. You give it a broad direction, and bundled agent skills—exploration, literature survey, experiment design, and paper writing—produce real artifacts: PDFs, Jupyter notebooks, figures, and reproducible run records. Everything stays in local workspace folders by default, with a provenance trail linking every figure and number back to the code and conversation that generated it.

The interesting bit

Instead of treating the LLM as a chatbot, the app treats it as a pipeline orchestrator. Each stage drops an inspectable, gradeable artifact into your filesystem—not just text, but runnable code, rendered LaTeX, and structured data. The built-in OpenCode sidecar makes it model-agnostic, while MCP connectors pull from arXiv, PubMed, Materials Project, and others without forcing you into a single provider’s ecosystem.

Key highlights

• Local-first by default: sessions, notebooks, run logs, and provenance live in local folders; nothing leaves your machine unless you explicitly connect a remote model or compute surface.
• Full provenance tracking: .openscience/provenance.jsonl links every artifact to its generating code, inputs, and conversation.
• Real notebooks and runs: native .ipynb support with Python/R kernels via bundled uv, plus append-only run logs and a SQLite run index.
• Scientific viewers out of the box: molecules, 3D meshes, genome tracks, FITS astronomy files, DOSCAR band structures, and standard office documents.
• Human-in-the-loop safety: command execution, file deletion, dependency installation, and remote connections require explicit approval in the UI.
• Multi-language UI: English, Chinese, Japanese, Spanish, German, French, and Korean.

Caveats

• Builds are not code-signed or notarized yet, so macOS Gatekeeper and Windows SmartScreen will complain.
• The project is explicitly a working MVP in active development; expect rough edges and shifting internals.
• Portuguese and Arabic are registered as interface languages but not yet selectable.

Verdict

Worth a spin if you want a local, model-agnostic lab notebook that actually produces files instead of chat bubbles. Skip it if you need code-signed installers or a finished product—the README itself bills this as a working MVP with unsigned builds.