Private Structure Mode, and Why We're Building a Free, Self-Hostable Chemistry Harness

Researcher accounts can now turn on Private Structure Mode — ZDR-routed AI requests and no default retention of chemical structures. It's also the first step toward a bigger goal: OpenMolClaw, a free, open-source, self-hostable chemistry agent harness, coming soon from the ChemIllusion team.

A single SMILES string, MOL block, or reaction scheme can disclose a proprietary lead compound, unpublished SAR data, a target-specific ligand, a synthetic route, or patentable subject matter. That's a different risk profile than a chat message. Generic "we don't train on your data" language, however true, doesn't say anything specific about how a structure itself is routed, stored, or looked up — and for research users, that's the part that actually matters.

This post covers two things: Private Structure Mode, a researcher-exclusive privacy posture that's live today on ChemIllusion, and OpenMolClaw, a free, open-source, local-first chemistry agent harness we're building — coming soon — for anyone who wants to keep their structures on their own machine entirely.

A SMILES string is not ordinary text

Chemical structures carry information density that plain text doesn't. A reaction SMILES with atom maps can describe an entire synthetic route. A MOL block from an unpublished project can be matched against public databases the moment it touches an external lookup service. For academic and industry researchers working on unpublished leads, regulated chemistry, or confidential collaborations, we think that deserves a narrower but stronger promise than a blanket privacy policy can make.

What Private Structure Mode does today

Private Structure Mode treats chemical structures as sensitive research content. When you enable it, supported AI requests containing structures are routed only to Zero Data Retention (ZDR) endpoints via OpenRouter, structure content is excluded from model training, external structure lookups (PubChem) are disabled by default, and raw structures are not saved to projects, chat history, session analytics, or logs unless you explicitly choose to save them.

Concretely, when the mode is on:

  • Eligible AI requests set OpenRouter's zdr: true, data_collection: "deny", and allow_fallbacks: false — no silent provider fallback.
  • If no ZDR-compatible endpoint is available, the request fails closed. Nothing is sent to a non-ZDR provider as a fallback.
  • External PubChem/PDB lookups by name or SMILES are disabled unless you explicitly enable lookup for that request.
  • Raw structures aren't written to project state or chat history by default — only if you explicitly save them.

It's exclusive to Researcher accounts on Premium or Professional plans, not teachers or students — teacher/student workflows lean on persistence, PubChem lookup, and saved history in ways that Private Structure Mode intentionally turns off, so it would work against those use cases rather than for them.

What we're not claiming

We'd rather under-promise here. Private Structure Mode does not mean no one can ever view your structures, that there is no chance of retention anywhere, or that every ChemIllusion workflow or every model/provider we use is ZDR. Structures still require transient processing at the provider endpoint to answer your request. ChemIllusion still retains non-content operational metadata (request time, feature used, token counts, cost, account identifiers) so we can run the product and bill correctly. And anything you explicitly save, export, or share can retain content — that's your choice to make, not something Private Structure Mode overrides.

How to enable it on chemillusion.com

If you're on a Premium or Professional plan:

  1. Go to Account → Privacy & Data.
  2. Under Your role, select Researcher.
  3. Toggle Enable Private Structure Mode on.

That's the whole flow. If you registered through a role-specific link and are already marked as a researcher, you'll see the toggle immediately without needing to change your role first.

The bigger goal: a free, self-hostable harness

Private Structure Mode is what we can offer as a hosted product. But the strongest privacy guarantee isn't a hosted setting at all — it's not sending your structures to a hosted service in the first place. That's the idea behind OpenMolClaw, a project we're building — coming soon — as a free, open-source, local-first chemistry agent harness: Ketcher for structure editing, RDKit for validation and rendering, and the same router-first tool-calling pattern that powers ChemIllusion's own chemistry agents, all running on your own machine.

We want OpenMolClaw to be the first genuinely open-source molecular agent harness that anyone — a grad student, an independent researcher, a small lab without an enterprise data agreement — can run locally, inspect, and extend, without needing a ChemIllusion account or sending anything to us at all. It's explicitly not a stripped-down demo of the hosted product: it's meant to be a complete, useful local tool in its own right, built and maintained by the ChemIllusion team.

It won't include ChemIllusion's hosted authoring workflows — Graphical Abstract preparation, CourseHouse, Journal Club, ChemLingo tutoring, video generation, LMS integration — those stay part of the hosted product. What it will include is the generic core: drawing and editing structures, RDKit-backed validation and rendering, a local workspace, and a router-first agent loop you can point at a model of your choosing.

Why explicit model routing matters

OpenMolClaw's model policy is the same philosophy as Private Structure Mode, just pushed further: no hidden vendor defaults, no silent fallback, and no guessed endpoints. You choose where your requests go — a fully local model on your own machine, or a hosted model you explicitly configure. When you want a hosted model without running your own inference, we recommend routing through OpenRouter with your own API key and requesting a ZDR-compatible model — exactly the same request-level controls (zdr, data_collection: deny, allow_fallbacks: false) that back Private Structure Mode on ChemIllusion. If a provider or model is named without an explicit endpoint, OpenMolClaw won't guess one — it fails closed instead.

OpenMolClaw's GitHub repository isn't public yet — we want the redaction, licensing, and documentation right before opening it up. When it ships, it'll be linked from the researcher pricing page and from this blog. If you want a hosted, zero-setup version of that same privacy posture in the meantime, that's exactly what Private Structure Mode is for today.

Try it

Researchers on Premium or Professional can turn on Private Structure Mode right now from Account → Privacy & Data. Everyone else can keep an eye on the researcher pricing page for when OpenMolClaw goes public.