Quick answer

Local installation alone does not establish a private chemistry workflow. Trace each operation: where inputs are processed, which remote services are called, what is retained, and which credentials are used. For example, SIRIUS combines local computation with integrated web services. This guide proposes a data-flow audit; it does not certify privacy or report a measured network audit.

“Installed on my laptop” describes a program’s location, not every destination of the data it handles. A private chemistry workflow needs a stage-by-stage account of files, requests, credentials and outputs. This guide helps researchers decide where unpublished structures, spectra and experiment tables may travel. It complements the environment-focused local deployment checklist with a data-flow audit.

Distinguish computation, download and submission

A local calculation reads inputs and computes on your machine or an approved local server. A remote API sends a request to another service; a hosted platform may additionally retain jobs, files or shared result pages. A model download retrieves software assets, while hosted inference submits research input. These are different transfer decisions, even when one Python function hides them.

Apply the distinction at the operation level. A local assistant can still send prompts to a hosted language model; a local MCP server can wrap an external API. Filenames, SMILES, spectra, parameter tables, error traces and retrieved identifiers can all reveal research activity. Inventory the actual configured components, not only their product names.

Classify the five example tools by the selected operation

Component/operation Local part supported by project documentation Boundary to inspect
RDKit parsing/descriptors Molecular manipulation and descriptor calculation in Python Enrichment wrappers, notebook hosting and assistant prompts may transmit the structures
ASE structure/optimizer with a selected calculator Local orchestration and atomistic objects The calculator may launch another program or use an external service
BoTorch candidate optimization PyTorch-based optimization/model calculations Objective evaluation, tracking and storage are supplied by the surrounding application
SIRIUS Local import, storage and several analysis stages Structural elucidation features use web services; inspect chosen features and subscription
GNPS2 web workflows Hosted mass-spectrometry analysis Uploaded files, tasks, result sharing and downstream destinations

These distinctions follow RDKit’s introduction, ASE calculators, BoTorch’s project description, SIRIUS technical documentation and the GNPS2 hub. They do not certify network behavior for an arbitrary installation. Code licences, model weights, data rights and hosted-service conditions remain separate review items.

SIRIUS: a desktop interface can be a service client

SIRIUS’s technical documentation distinguishes local stages from web-service features such as CSI:FingerID and CANOPUS. It lists transferred scientific inputs, including spectra and fragmentation trees, and describes authentication/licence endpoints and application servers. A local REST API can therefore expose operations that subsequently call remote services; “localhost” alone does not settle where computation occurs.

Read the operator’s current retention/security statements and the applicable subscription terms before using private data. Statements about transient processing or encrypted transport do not remove the need to authorize transmission. The client repository’s AGPL licence does not grant unrestricted use of every hosted feature. This article does not audit the operator’s infrastructure or provide legal conclusions.

GNPS2: uploaded analysis is an online branch

The GNPS2 hub describes a web-based platform and exposes public tasks and multiple services. Those facts do not establish that every private job has identical sharing, retention or deletion settings. Do not transfer classic GNPS/MassIVE publication behavior to a GNPS2 workflow without checking that workflow.

Before submitting unpublished spectra, identify the precise workflow, upload host, account permissions, task visibility, result links, retention/deletion policy and any downstream calls. If those facts are unavailable, mark the hosted branch pending authorization and keep private fixtures out of it. Using local alternatives can change the available capabilities; do not promise that a local spectral preprocessing step reproduces all hosted analyses.

Conduct a bounded network audit

Use public or synthetic fixtures first. The procedure below is a proposed operational review; no network audit was executed for this article. Code status is not applicable because no portable command can prove privacy across all operating systems and deployment contexts.

  1. Draw the intended flow: input files → local preprocessing → optional external analysis → results → assistant/report. Annotate each edge with data classes and destination ownership.
  2. Separate installation and first-run asset acquisition from routine processing. Record package indexes, model registries, licence checks, telemetry, update requests and any externally mounted/cloud-synced output folders.
  3. Review configuration, wrappers, calculators, plugins, background workers and child processes. Enumerate read/write paths, subprocess permissions, approved domains and credential scopes. Store secrets outside prompts and shared logs.
  4. Observe DNS and connections for the actual process tree using institution-approved OS, container or proxy facilities. Keep timestamp, executable, destination, purpose and a redacted account of the payload class. TLS destination visibility alone does not reveal the request body; do not claim otherwise or disable certificate checks to obtain a log.
  5. After approved dependencies/assets are available, exercise the intended local branch with outbound access denied in the execution environment. Record failures, retry behavior and delayed uploads; repeat any online branch only with its separately approved destinations and public fixtures.
  6. Review application logs, model caches, notebooks, backups and result-sharing settings. A single run with no observed connection supports only that bounded observation, not a universal no-transfer guarantee.

Turn observations into an approval record

For each stage, record the data owner/classification, exact inputs and outputs, software/model revisions, local read/write scope, endpoints, payload class, purpose, retention basis, responsible reviewer and evidence timestamp. Distinguish “declared,” “observed,” “blocked” and “unknown.” An allowlist is an enforcement rule only when applied by an actual runtime or network boundary; a document alone cannot prevent a request.

A failed offline run can expose an unexpected dependency. Do not silently re-enable the internet on private data. Review whether the request is an asset fetch, licence verification or scientific-data submission, and whether a reduced local workflow still meets the task. Removing obvious names is not necessarily anonymization: molecular structures and unique spectra can themselves be confidential.

Recipe and CARS: describe permissions before claiming enforcement

The existing local assistant and cleaning Recipes are useful starting points, but inspect their exact components: optional PubChem enrichment or a hosted assistant changes the transfer map. A catalog Recipe does not inherently provide an execution sandbox or a confidentiality guarantee.

CARS is currently an experimental Community Draft 0.1. Its prose specification proposes security/network declarations. The schema notes limit the available reference validator to L0 structure checks, which do not enforce permissions, audit dependencies, resolve adapters or validate science. Permission enforcement, local execution and recorded egress would require a concrete runtime and evidence. Those are potential integration tasks, not capabilities established by passing a manifest validator. No new Recipe, CARS adapter or privacy certification is created here.

Related ChemAI Atlas resources and workflows

RDKit · ASE · BoTorch · SIRIUS · GNPS

Local Deployment Checklist for Chemistry AI · MS/MS Annotation Workflow: matchms, MS2DeepScore, SIRIUS and GNPS2 · What Is CARS? A Guide to the Chemistry AI Recipe Specification · Reading Code, Weight and Data Licences Separately

Build a Local Chemistry Assistant with Ollama + RDKit · Chemical Data Cleaning Stack

CARS

Sources and evidence status

Sources inspected on 2026-10-18. Primary documentation supports capability/API statements. Audit protocols are editorial proposals; no scientific validation or measured performance is reported.