Managed OpenScience hosting: own the open-source Claude Science workbench

OpenScience is the open-source AI workbench people mean when they say “Claude for science” without the black box.

You give it a research goal. It reads the literature (arXiv, Semantic Scholar, OpenAlex), forms a hypothesis, writes and runs code, hits scientific databases (UniProt, PDB, ChEMBL, PubChem, …), and writes up what it found. Specialist agents for research, ml, biology, and physics. Two hundred ninety-plus skills. A real browser workspace — file tree, editor, terminal, inline molecules and plots.

What it is not: a shared SaaS notebook that holds your keys and meters your model spend.

The problem with “just run it locally”

npm install -g @synsci/openscience && openscience works — until it doesn’t.

• Your laptop sleeps. The literature sweep dies mid-session.
• A VPS is a snowflake: Node, ports, reverse proxy, backups, “who has sudo.”
• The workspace binds loopback only by design. Fine on a desktop; painful when you want a stable HTTPS URL from another machine.
• Credentials land in XDG dirs under $HOME. Snapshot carelessly and you ship API keys with the template.

Scientists and ML engineers want the agent always on, isolated, and theirs — keys inside the box, not proxy-marked-up through a vendor.

What managed OpenScience on jurniti actually is

One Firecracker microVM per tenant. Same hardware isolation primitive AWS uses for Lambda.

Pricing is flat per VM (Starter / Pro / Max). OpenScience’s research workbench is Pro floor — the agent + workspace need headroom Starter’s 2 GiB doesn’t have.

Who this is for

• ML researchers who want a co-scientist that can lit-review, scaffold training/eval, and write methods — without babysitting a laptop process.
• Comp-bio / structural biology teams who need UniProt/PDB/ChEMBL in the tool loop and a workspace that doesn’t share a kernel with strangers.
• Labs and indie PIs who refuse to put grant-funded keys into a multi-tenant chat product.

If you only want a chat UI and a free tier, this is the wrong product. If you want ownership, keep reading.

How provision works (about three minutes)

1. Pick OpenScience at checkout (Pro or Max).
2. Stripe confirms → fleet assigns a host → microVM boots the pre-baked rootfs.
3. Open the workspace URL. Add a provider key (openscience keys add or the Credentials panel).
4. Give it a goal. Use ml or biology when the default research agent isn’t the right specialist.

No founder SSH. No “we’ll set it up for you on Tuesday.”

Templates: don’t start from zero

Two high-signal public templates (when founder-gated live):

1. AI/ML Research Co-Scientist — lit → hypothesis → train/eval scaffold → write-up.
2. Comp-Bio / Structural Biology Research Agent — target/structure queries against real scientific DBs → grounded brief.

Fork → BYOK → one concrete research goal. That’s the aha.

OpenScience vs “just use ChatGPT with papers”

Chat products summarize PDFs you paste. OpenScience runs the loop: tools, code, databases, multi-step sessions, artifacts on disk. You can export, snapshot, and take the upstream agent elsewhere — Apache-2.0, no jurniti fork required.

FAQ

Do I need Atlas (Synthetic Sciences’ managed wallet)?
No. BYOK is first-class. Atlas is optional.

Can I use it headless?
Yes — openscience serve is the headless path. On jurniti the portal reverse-proxies the workspace so you get HTTPS without fighting loopback bind.

Is my data used to train someone else’s model?
Not by jurniti. Your keys, your provider, your persist volume. Templates scrub auth and sessions before public share.

Why not E2B / a generic sandbox?
Those are code-execution sandboxes. This is a managed agent harness with a full research workspace, billing, templates, and fleet that only rents boxes when prepaid cash covers them.

Get started

Managed OpenScience on jurniti — or the free 7-Day OpenScience Setup if you want the ownership arc first.

Build the science. Own the agent.