ai4s-research/open-science

What it solves

It provides a local-first, model-agnostic workbench for the entire scientific research loop. It eliminates the "black box" nature of AI research by ensuring every artifact—such as figures, tables, and reports—is auditable and traces back to the exact code, inputs, and model conversations that produced it.

How it works

Built with Tauri and React, the desktop application uses a bundled OpenCode sidecar to interact with various LLM providers. It employs a system of "skills" (specialized agentic workflows) to handle tasks like literature surveys and experiment design. The system maintains a local provenance log (provenance.jsonl) to track file versions and links. It also integrates with MCP (Model Context Protocol) servers to connect to scientific databases (e.g., arXiv, PubMed) and can control a user's own Chrome browser to harvest live web data.

Who it’s for

Researchers and scientists who need an autonomous AI assistant to help with exploration, literature review, coding experiments, and paper writing while maintaining strict data privacy and reproducibility.

Highlights

• End-to-End Research Loop: Automates the flow from broad exploration to a finished research paper.
• Full Provenance: Every output is linked to its generating script and input data for complete auditability.
• Local-First Privacy: Sessions, data, and run records are stored locally by default.
• Browser Control: Agents can drive the user's actual Chrome profile to access the web with existing logins.
• Remote Access: Includes a token-authenticated gateway to monitor or trigger runs from a phone or LAN browser.
• Science Connectors: Built-in MCP integrations for biomedical, materials, and economic databases.