About this template
From a pile of PDFs to a methods draft
If you start ML research with forty open tabs and an empty repo, this OpenScience template hands you the co-scientist loop already wired: literature, hypothesis, experiment scaffold, methods write-up.
Files come out. Not a chat monologue.
What the ML co-scientist loop does
Two skills carry the work.
ml-lit-loop runs the literature side — survey, shortlist, notes you can return to. ml-experiment-scaffold turns that shortlist into one concrete experiment: a train/eval scaffold plus the methods section that describes it. Output lands as files under the workspace layout, so tomorrow's session picks up where this one stopped.
You still decide which experiment is worth running.
What lands in your fork
- The OpenScience ml agent, oriented for research co-scientist work
ml-lit-loop— the literature loopml-experiment-scaffold— experiment scaffold and methods draft- A layout for lit notes, the scaffold, and the write-up
- Browser workspace entry via Open OpenScience
What you supply (BYOK)
One model key gets you moving.
ANTHROPIC_API_KEY is the documented path; OPENAI_API_KEY, GEMINI_API_KEY and OPENROUTER_API_KEY are read too. Export it in the VM shell:
export ANTHROPIC_API_KEY=sk-ant-...
Keys stay on your microVM. Nothing is baked into the template.
Your first five minutes
Fork, open the VM, click Open OpenScience.
Then give the ml agent one full instruction:
Survey diffusion models for medical imaging. Propose one experiment. Scaffold train/eval. Draft a methods section.
You should end up with a paper shortlist, one concrete experiment plan, a code scaffold, and a methods draft you can edit — not a transcript.
Honest limits
This is a research desk, not a training cluster.
There is no GPU in the box and no managed fleet behind it: the scaffold gets written here, the heavy training runs wherever you run training. Nobody else holds your API keys. It is ML-shaped, so for pure biology take the comp-bio OpenScience template instead. Two skills and a layout is the honest surface area — the depth comes from what you ask it.
Model usage is yours. 30-day money-back on your first jurniti purchase.
Fork it. Open OpenScience. Run one survey-to-scaffold pass.