Atom Privacy Note

Beta build 0.1.4, 2026-11-01 (the product was called Corroborate Studio before 0.1.4).

This note is from Atom Industries [legal entity suffix to confirm] ("we", "us"), the maker of Atom (the "Software").

In short: Atom runs on your computer and sends us nothing. When you use the agent, it sends data to the AI provider of the model you chose (Anthropic, unless you choose another) under your own API key, and that key stays on your machine. Web search on another provider's model also goes to Anthropic, through your Claude key.

1. No telemetry. The Software has no analytics, usage tracking, crash reporting, advertising or automatic update checks, and needs no account with us. It never sends us your models, specs, results, logs, settings or key. The app window loads nothing from the internet: it talks only to the Atom backend on your own computer (127.0.0.1). A web link you click opens in your own browser.

2. What goes to AI providers, and when. Only when you use the agent features (the agent panel, building a model from a description, Fix and reverify, and Test connection in Settings) does the backend send requests to an AI provider, with your API key for it. By default that is Anthropic's Claude API. A chat thread runs on the provider of the model you pick for it, which can also be OpenAI, xAI, DeepSeek, Meta, Google (the Gemini API), Mistral AI, OpenRouter, an Ollama server or an OpenAI- or Anthropic-compatible endpoint you configure; building a model and Fix and reverify always use Claude. A request holds your messages, the Software's instructions to the model, and what the agent reads through its tools: files in your workspace, model trees, properties and parameters, evaluated results, verification results and solver logs. It also carries the standard request headers of the provider's Python SDK (Anthropic's, or OpenAI's for the other providers), such as the SDK version, your operating system and the Python version. Each provider handles this data under your agreement with it and its privacy policy, not under this note. Their terms differ: DeepSeek, for example, stores data in the People's Republic of China and does not promise not to train on it, and Google uses content sent on the free tier of the Gemini API to improve its products, which is why Settings asks you to confirm that billing is on before a Gemini model runs. Besides chat requests, the backend asks a provider for its list of models (no messages, no tokens) when you save its key in Settings, press Test connection or Refresh, or start a chat on a model it has not listed yet; for a key you saved in Settings it also does so when the app starts. A key that is only in your environment is not sent to its provider until you do one of those things. The Ollama server at the URL in Settings (by default on this computer) is asked for its models when the app starts, without a key unless you saved one for that URL. If you never use the agent features or a provider's settings, nothing is sent to that provider. Never send export-controlled data (section 6 of the license agreement).

Web search and web fetch. In chat threads and model builds on a Claude model, the agent can use Anthropic's web search and web fetch tools unless you turn web search off in Settings; Fix and reverify never uses them. In a chat thread on another provider's model, Atom runs the same two tools itself through your Claude key: each search or page read is one request to Claude Haiku 5.5 on Anthropic's Claude API, so the query (or the address of the page and what to quote from it) goes to Anthropic, with the domains Settings limits web search to if you list any, and the results then go on to that chat's provider with the rest of the conversation. Without a Claude key those models answer without web search. Anthropic runs the search with its search provider and fetches the page from the website; the website sees a request from Anthropic, not from your computer, and may keep the address and request details under its own policies. The results and the page text come back into the conversation and are stored with the thread on your computer. The agent is instructed not to put your simulation numbers, run ids or file contents in a query, and it can fetch only an address that you typed in the thread or that a search returned (or the page that address redirects to), never one found in your files. Anthropic charges 10 USD per 1,000 searches to your key, and on another provider's model also the Claude Haiku 5.5 tokens of each search and page read; the agent's cost includes both. Settings can limit search and fetch to the domains you list, and turning web search off stops it on every model.

3. Your API keys stay on your machine. The installed app saves each key you enter in Settings to a file in your Windows profile, %APPDATA%\Atom Industries\Atom\credentials.env, or reads it from an environment variable (ANTHROPIC_API_KEY for Claude, OPENAI_API_KEY for OpenAI, and one such variable per provider). That file is plain text, protected by your Windows account's file permissions, not encrypted. It also records which host a key for an endpoint you configured may be sent to, and which data-use notices you confirmed; it holds no other data. Each key goes only to its own provider, to authenticate each API request, and a key for an endpoint you configured goes only to the host you saved or tested it for; the app never shows a key again and masks keys in its log. On a roaming Windows profile the %APPDATA% folder can be copied to your organization's servers; to keep the key on this machine only, add "extraArgs": ["--env-file", "<a local folder>\\credentials.env"] to the app's settings.json.

4. Data on your computer. Your workspace (Documents\Atom unless you choose another folder; on an install that was upgraded from Corroborate Studio, the first start of 0.1.5 renames Documents\Corroborate Studio to Documents\Atom, and the app's own folder in a workspace from .corroborate-studio to .atom (.atom-industries when the workspace is your home folder), on this computer only and only when nothing else uses them; until then the old names stay in use), your models, verification runs and reports stay where you keep them. The app keeps its settings in %APPDATA%\Atom Industries\Atom, and its logs and window storage in %LOCALAPPDATA%\Atom Industries\Atom. The first time Atom starts after it was installed over Corroborate Studio, it copies settings.json, window-state.json and credentials.env from %APPDATA%\Corroborate Studio into its own folder, on this computer only. It leaves the old files as they were and adds one short note, moved-to-atom.txt, to the old folder, so the copy happens only once: deleting Atom's folder later does not bring the old files back. The backend log (logs\backend.log) masks keys, tokens, passwords, email addresses, account and machine names and file paths before it is written. Nothing is sent anywhere; you decide whether to share a log with us for support.

5. COMSOL. The live worker runs your own COMSOL installation, which contacts your COMSOL license server the way that installation is set up to. That traffic is between you and COMSOL.

6. Removing your data. Uninstalling removes the program and keeps %APPDATA%\Atom Industries\Atom (with credentials.env), %LOCALAPPDATA%\Atom Industries\Atom and your workspace, and after an upgrade also %APPDATA%\Corroborate Studio (with the old credentials.env) and %LOCALAPPDATA%\Corroborate Studio. Delete those folders to remove everything. To stop a key from working anywhere, revoke it in your account with its provider.

7. Changes and contact. If a later build changes what is sent or where, its privacy note will say so. Contact: [CONTACT ADDRESS TO BE ADDED BEFORE RELEASE]