A new desktop app called Open Science just launched as an open-source, local-first alternative to Claude Science, and it debuted on Hacker News where it climbed to a score of 160. The project, built by the ai4s-research team and posted as a Show HN, reframes AI-assisted research around a workbench instead of a chat box. According to Hacker News, the pitch is simple: one prompt should drive a complete, traceable analysis where every number, figure, and report links back to the code that made it.
The headline idea here is provenance. Most AI research tools hand you text and hope you trust it. Open Science does the opposite. Click a figure and it shows you the script behind it, the data it used, the environment it ran in, and the conversation that produced it.
What it actually does
The workflow runs through four stages: plan, approve, execute, review. You stay in the loop the whole way. Here’s what stands out from the launch details:
- Traceable artifacts. Every figure, table, notebook, and report links back to its code, data, and environment. A file called provenance.jsonl records each version so you can recover any past state.
- Local-first by default. Your data, code execution, and session history stay on your machine. The Settings panel spells out in plain language exactly what gets sent to your model provider and what never leaves.
- Model-agnostic. Bring your own key through OpenRouter, OpenAI-compatible endpoints, Anthropic, or a local Ollama model. Around 150 providers are supported, and a free built-in model works with zero setup.
- A real Jupyter kernel. The agent drives a live Python kernel. Cells and figures appear beside the chat, and you can sweep parameters across experiments while keeping a persistent session.
- A traceability reviewer. This one is the sharpest feature. It resolves citations against Crossref, arXiv, and PubMed, flags numbers with no traceable source, and checks whether figures actually match the code that generated them.
Built for reproducibility
The whole thing runs on a bundled OpenCode agent runtime, a single-binary sidecar the app pins and manages. The UI never talks to a model directly. It goes through a thin SDK, which keeps skills, MCP servers, and model providers pluggable.
Open Science ships with a set of scientific skills out of the box: research-explorer, literature-survey, experiment-suite, paper-writer, integrity-auditor, and more. It also includes one-click connectors for literature sources like Semantic Scholar and bioRxiv, plus biomedical tools like ClinicalTrials.gov. You can wire in any MCP server you want.
Why this matters: reproducibility is a real crisis in computational research. Papers get published with figures nobody can regenerate and numbers nobody can trace. A tool that audits its own output for citation gaps and figure-to-code consistency is aiming straight at that problem.
Who can use it, and the catches
This is early, and it shows. There’s no polished installer waiting for you. According to the Hacker News post, you build the desktop app from source. That means cloning the GitHub repo and running through pnpm, with a few scripts to fetch the pinned runtime, the uv Python environment manager, and the skills pack.
Prerequisites are real: Node.js 20 or higher, pnpm 9, and the Rust toolchain for Tauri. It targets macOS or Windows. On first launch it starts the runtime automatically and works with the free model, so you can try it before plugging in your own provider.
A few honest limitations worth flagging. Manual approval is the default, so the agent asks permission before running commands or writing files. That’s a safety win, but it also means this isn’t a fire-and-forget tool. And the build-from-source requirement puts it out of reach for non-technical researchers for now.
The roadmap starts with AI for Science and expands toward materials, chemistry, biology, medicine, and engineering. If the team can turn the source build into a clean installer, this could become the reproducibility layer a lot of computational scientists have been asking for. Full setup instructions and screenshots are available at the original source.