Ask any researcher what actually eats their time, and it's rarely the thinking. It's the grind between the idea and the conclusion: reading dozens of papers, setting up environments, writing code, running experiments, cleaning data, and writing it all up.
OpenScience (3.4k stars, Apache-2.0, TypeScript) wants to hand that whole chain to an AI. Its pitch: give it a goal, and it reads the literature, writes and runs code, runs the experiments, and writes up what it found.
It runs the full loop, not just chat
OpenScience isn't "ChatGPT for research." It runs a complete research loop the way a capable collaborator would:
read the papers that matter → form a hypothesis → write code → run experiments on your compute → query scientific databases → write up the result.
All in one continuous session, in a browser workspace with a file tree, editor, terminal, and inline rendering for molecules, structures, genomes, and plots.
The parts that signal it's serious
- 312 bundled skills — real research workflows, not stuffed prompts: training (DeepSpeed, PEFT, TRL), evaluation, dataset work, molecular and clinical biology, cheminformatics, papers and LaTeX, figures.
- 42 scientific database connectors — UniProt, PDB, Ensembl, ChEMBL, PubChem, arXiv, OpenAlex, Semantic Scholar. This fixes a real pain point: an AI that can actually look things up instead of hallucinating citations.
- Model-agnostic — frontier or open-weight models, your own keys, or local models. Keys stay on your machine; requests go straight to the provider.
- Governed capabilities — a three-tier permission model (Approve for me / Ask for approval / Full access) with commands running under macOS Seatbelt or Linux bubblewrap containment.
An honest boundary
The README is refreshingly explicit: the permission system is awareness, not isolation, and the sandbox is not a full jail — use a container or VM for hostile code. That kind of honesty about limits is rare in AI projects, and it's exactly what you want in a tool whose output might influence experimental conclusions.
Who it's for
Researchers in ML, biology, physics, and chemistry who want to accelerate the "literature → experiment → write-up" pipeline — and teams who want to run it on local models with data staying on their own hardware.
It's not a paper generator. It's a collaborator for the grunt work. The judgment of whether a hypothesis is right, and whether a result is trustworthy, is still yours.
I've localized the README and core docs to Chinese: https://github.com/yangshun2005/openscience-cn
If you find this project useful, a star on the original repo supports the author's ongoing maintenance.
Top comments (0)