ai4s-research/open-science

Open Science Desktop is a desktop application, built for macOS, Windows, and Linux, that aims to run an entire research project inside one continuous, traceable session. It covers the whole loop from an initial idea to a finished paper: exploring the topic, surveying existing literature, forming a hypothesis, writing experiment code, analyzing results, producing figures, and writing up the findings. The project describes itself as an open-source alternative to Claude Science and similar AI-for-science tools, and it is built with Tauri, MCP, and agent skills. A bundled agent chains together specialist skills, one for exploring, one for surveying literature, one for experimenting, and one for writing, so each stage produces a real file you can inspect, such as a markdown document or a notebook, rather than just a chat message. Every figure, table, report, and run result links back to the exact code, inputs, environment, and model output that created it, so results can be traced and checked. The application is local-first: sessions, data, notebooks, and run records are stored in folders on your own machine, and nothing is sent elsewhere by default. It is also model-agnostic, meaning the interface talks to a bundled runtime called OpenCode, and you can choose which underlying AI model powers it. Runs can happen locally or on remote resources such as SSH, Slurm clusters, or Modal, and each is captured as a reproducible record rather than scrollback text in a terminal. A built-in gateway, off by default, lets you check on a running analysis from a browser on your phone or local network once you turn it on, without your API keys leaving your machine. The agent can also control your own Chrome browser, using your existing logins, to read live web pages, or use a separate private browser if you prefer. The project is licensed under MIT and reports ranking first on an external benchmark for autonomous research agents called ResearchClawBench.