Open Science Desktop

This is a desktop research workbench for scientists who want AI help without sending project files to a hosted service. It works with any model provider you bring, and it's an open alternative to Claude Science. Runs on macOS, Windows and Linux.

Every request follows the same path. The agent states its sources, steps, expected outputs and risky actions first. Then it executes against your local workspace. Then you review the evidence.

• Literature to reports: search papers, parse sources, draft manuscripts, and keep citation limits visible.
• Code-backed analysis: run Python and notebooks locally, and keep the scripts behind each result.
• Artifact provenance: figures, tables, PDFs and notebooks keep their generating steps, inputs and versions.
• Approval gates: command execution, dependency installs, deletion and network access need your explicit consent.
• Auditable manuscripts: resolve citations and flag unsourced claims before export.

Open a figure and you can see the script, input data, model context and conversation that produced it. A single prompt can simulate data, fit a model, save a figure and write a report with that evidence trail intact.

Project files stay local by default, and the bundled runtime is isolated from the rest of your system. The app is built with Tauri and React, uses MCP for tool connections, and builds on OpenCode. Provenance is stored in SQLite and JSONL, so it's inspectable. Installers cover macOS 13+, Windows 10/11 and Linux deb/rpm packages.