Open Science Desktop Released

Open Science Desktop, an open-source desktop tool that transparently records the Provenance and execution environment of figures, tables, reports, and other outputs generated during AI research, has been released. By linking generated code, input files, and execution environment information scattered across conversation logs to the artifacts and storing them, it helps easily provide evidence during paper reviews or experiment reproduction.

Key Features

• Local-first architecture: Sessions, data, provenance records, and execution logs are all stored on the user's computer, with no data leakage to external sources by default.
• Artifact tracking: Opening a figure or table file instantly reveals the script code, input data (CSV, etc.), execution environment, and conversation logs that generated it via the Artifact Inspector.
• Execution record management: Remote execution histories via SSH, Slurm, Modal, etc., as well as local executions, are stored in a reproducible format and can be searched and filtered through a global SQLite index.
• Research Skill-based workflow: Skills for each research stage, such as literature review, experiment design, paper writing, and integrity audits, are provided, with each stage generating gradable artifacts.
• Browser control: Web pages can be read via the Accessibility Tree while maintaining the user's actual Chrome profile, facilitating data collection from sites requiring login credentials.

Tech Stack and Caveats

Developed with Tauri 2 and React, model execution supports various LLMs via the OpenCode sidecar. Builds for macOS, Windows, and Linux are provided, but the project is currently in the MVP (Minimum Viable Product) stage, and Windows/Linux builds are unsigned. Additionally, caution is needed as it may be confused with similarly named projects (OpenScience by Synthetic Sciences).