OpenScience

OpenScience is an open-source AI workbench for scientific research that strings together the full experimental loop — from a stated research goal, through literature review, hypothesis generation, code and data analysis, to a drafted paper. Rather than locking you into one provider's model, it's model-agnostic: you point it at whatever LLM you want to run, and it orchestrates a library of 250+ research "skills" that are themselves editable rather than baked into a black box.

It's built by Synthetic Sciences (YC W26) and launched July 5, 2026 as a deliberate open alternative to the closed scientific-agent systems shipping from big labs — most notably Google's Co-Scientist. The pitch is transparency and control: because the skills and orchestration are open, a researcher can inspect exactly what a step does, fork it, and swap in their own tools or datasets. That matters in science, where reproducibility and knowing why an agent reached a conclusion is often the whole point.

This is a tool for technical researchers and research engineers, not a polished consumer app. Being GitHub-distributed and self-hosted, it assumes you're comfortable wiring up API keys, running code locally, and picking your own models. In exchange you get no per-seat subscription, no vendor lock-in, and a skill library you can extend — a genuinely different bargain from the hosted research assistants like Elicit, Consensus, or Google's Co-Scientist.