open-science: a local-first AI research workbench that automates the scientific loop from literature survey to paper writing

What it solves

Open Science Desktop provides a unified, local-first workbench for the entire scientific research loop. It eliminates the fragmented workflow of switching between chat interfaces, notebooks, and file managers, while ensuring that every result—figures, tables, and reports—is auditable and traceable back to the exact code and data that produced it.

How it works

The platform is a model-agnostic desktop application built with Tauri and React. It uses a bundled OpenCode sidecar to interact with various LLM providers. The system organizes research into sessions and workspaces where autonomous agents (via the ai4s-agent) execute a chain of specialized "skills" (such as literature survey and experiment design). These agents produce real files and artifacts rather than just text replies, and a provenance system tracks the history of every file version and edit.

Who it’s for

It is designed for researchers and scientists who need an AI-powered environment that maintains strict data privacy, supports reproducible research, and integrates deeply with scientific tools like Jupyter notebooks and various academic databases.

Highlights

• End-to-End Research Loop: Automates stages from broad exploration and literature surveys to experiment code, analysis, and final paper writing.
• Full Traceability: Links figures, reports, and tables directly to the exact model output, inputs, and code used to generate them.
• Local-First Privacy: Keeps sessions, data, and provenance records on the user's local machine by default.
• Scientific Connectors: Includes one-click MCP connectors for arXiv, PubMed, Crossref, Semantic Scholar, and other specialized scientific databases.
• Reproducible Runs: Captures local, SSH/Slurm, and Modal runs as formal records rather than simple terminal logs.
• Extensible Architecture: Supports pluggable agent skills, MCP servers, and a model-agnostic SDK.