SandboxAQ connects scientific AI models to chemistry research workflows

SandboxAQ builds quantitative AI for scientific problems. Its documented chemistry tools let researchers request model calculations through an AI assistant, but the useful output remains a scientific result with defined inputs and limitations. This public-source blueprint proposes an internal evaluation; it does not report laboratory or product testing.

01 / ProductQuantitative models supply the calculation behind the conversation

SandboxAQ describes its approach as Large Quantitative Models, combining scientific data, physics and chemistry with AI. Its broader positioning spans several industries, but this blueprint concentrates on the publicly documented molecular and materials tools. That gives readers an inspectable entry point rather than treating every company research programme as a self-service product.

AQCat Adsorption Spin wraps a machine learning potential in a workflow that selects a bulk structure, creates a surface, places an adsorbate and relaxes geometries. Its job is catalyst pre-screening. The returned energies concern the specified material, facet and adsorbate; a fluent explanation from the assistant does not widen that scientific scope.

AQPotency instead accepts protein and ligand identifiers to produce relative potency rankings with uncertainty and applicability signals. Its documentation explicitly distinguishes ranking from reliable absolute potency prediction. Treating the two products as interchangeable chemistry chatbots would lose the most important distinction: they answer different kinds of computational question.

02 / AudienceThe right reader already owns a scientific decision

A computational chemistry team may want to reduce a candidate set before committing more expensive simulation or experimental capacity. SandboxAQ can be relevant when a documented model covers that particular screening task. The scientist still decides how the inputs represent the physical problem and what evidence would justify advancing a candidate.

A software team building research tools has a different question: can it connect a conversational interface to a reproducible calculation? The MCP documentation describes OAuth authentication and managed model execution. This is useful integration plumbing, but access to a tool is separate from permission to package it as a service for outside customers.

For a wider molecular design environment, the Schrödinger blueprint provides an adjacent comparison. For the language-model interface, the Anthropic blueprint covers another layer of the stack. Neither comparison establishes equivalent chemistry accuracy. The practical choice concerns where the calculation, research context and human judgment should reside.

03 / WorkflowA proposed catalyst evaluation starts with a traceable calculation

Consider an internal research group comparing adsorption on a small, already understood family of surfaces. Begin with a reference question for which the team has accepted results. Record the material identifier, surface facet, adsorbate and intended interpretation before invoking any assistant. This proposed exercise tests whether the service can support that team's next research decision; it is not a new scientific result.

Create an account through the documented onboarding route and connect a supported MCP client. Account registration, payment setup and the client subscription are separate considerations. Verify the authenticated tool list before uploading proprietary structures. A successful chat login alone does not show that the chemistry service is connected.

Ask explicitly for the AQCat tool and preserve the submitted parameters and returned job information. SandboxAQ's documentation warns that an LLM can answer from its own prior knowledge instead of invoking the intended tool. The evaluation should therefore reject an answer that lacks an actual calculation record, however plausible its wording appears.

Review every returned placement's convergence flag before comparing energy values. Exclude or investigate unconverged results rather than allowing the assistant to select the most attractive number. Then compare the accepted results with the team's reference under matching assumptions. A discrepancy can come from the input structure or simulation setup as well as the learned potential.

For an AQPotency evaluation, replace the surface experiment with a controlled molecular ranking exercise. Keep the protein accession and compound identifiers attached to each result. Review the uncertainty and training-domain indicators alongside the ordering. Reserve an independently selected set for downstream evaluation so that manual selection does not quietly make the model appear more useful.

Finish by recording the decision that the calculation changed. Perhaps the team advances fewer candidates, identifies an out-of-domain region, or discovers that its question needs a different method. Each is useful information. Counting generated answers or screened molecules would miss whether the scientific work actually became more discriminating.

04 / PricingUsage charges and enterprise access sit on different layers

The current pricing page lists AQCat at US$1 per Relaxation Unit: one bulk–facet–adsorbate workflow, with compute and orchestration included. AQPotency is billed per protein–ligand pair, with a fresh tier schedule for each call. Small separate calls and a batched call can therefore have different costs even when they evaluate the same total number of pairs.

The Individual tier has no monthly platform fee and is positioned for individual scientists, tool builders, evaluation and prototyping. Enterprise lists a US$50,000 annual platform fee per organisation or department, with separate negotiated terms. Production R&D and automated pipelines are placed in that enterprise column; a free platform fee should not be read as an unrestricted production entitlement.

The Master Customer Agreement frames use around internal business purposes and restricts service-bureau use and publication of benchmarks. A commercial external research assistant therefore needs explicit contractual coverage. The proposed evaluation here is internal, and any public performance publication should be agreed separately. Include the chosen LLM client's costs and downstream experimental work in the project budget.

05 / DistinctionsThe interface can expose scientific uncertainty instead of hiding it

One useful distinction is the separation between conversational orchestration and a specialised model result. The assistant helps express a task and interpret output, while the computational service provides structured fields. That makes it possible to demand provenance at the tool boundary instead of accepting an unsupported paragraph as the answer.

AQCat's convergence information and AQPotency's applicability signals support different rejection decisions. A failed relaxation is not merely a low-confidence potency estimate, and a ligand outside the useful training domain is not a missing server response. Preserve those distinctions in any internal report so a scientific reviewer can decide what to rerun and what to discard.

Our editorial assessment is that SandboxAQ is most interesting where it makes a narrow computational method easier to reach without erasing the method's boundaries. Its company-level marketing is broader than the specific evidence available from these tool manuals. The manuals are the better basis for selecting the first practical project.

06 / QuestionsResolve model fit and operational terms before scaling

Ask which model version produced a result and how upgrades affect repeatability. A candidate ranking may be used weeks after its initial computation, and the team needs to know whether it can reproduce that calculation. Save the input set, result fields and version information available from the service rather than relying on a chat transcript alone.

Latency statements differ across the general MCP page and the more specific AQCat workflow description. This blueprint therefore promises no completion time. A team with a scheduled laboratory queue should establish a measured operating envelope and obtain the service commitment appropriate to its tier, including what happens when a job fails or must be repeated.

Finally, confirm the permitted data and output uses for proprietary molecular libraries. Ownership, retention and integration permissions belong in the actual agreement. A model with useful scientific coverage can still be the wrong operational fit if the organisation cannot submit the required inputs or retain the evidence it needs.

07 / DecisionChoose a bounded scientific question before choosing an AI interface

SandboxAQ earns consideration as a scientific AI company because it exposes concrete quantitative tools with identifiable inputs, outputs and limitations. Start with the calculation you need to improve, then test whether the conversational route preserves scientific traceability. Expanding to unattended research pipelines should follow a successful internal evaluation and an appropriate enterprise agreement.