Data and AI

Last updated: 12 October 2026

What ArchX sends where, in plain words. The legal details are in the Privacy Notice and the Data Processing Terms.

In short

• Your files stay on your device. ArchX works on the files in your project folders on your computer. TYTOS does not receive them.
• The AI work goes directly to your AI provider. What the agent reads and writes for a task goes from your device to the provider you set up in the app: through your ChatGPT account to OpenAI, or with an API key to the provider of that key. It does not pass through TYTOS.
• TYTOS receives only what it needs for your licence: your licence email address, a random installation ID and details of your device. It keeps no usage counts for practices in the United Kingdom.
• Support is the exception you control: a report reaches TYTOS only when you send it, with the attachments you choose.
• For personal data of your clients, use a business plan with your provider’s data processing terms and without training (see the last section).

What stays on your device

• project files, drawings, documents and the results ArchX creates;
• your conversations with the agent;
• API keys, only in the operating system’s key store: the Keychain on a Mac, the Credential Manager on Windows;
• the sign-in details of your ChatGPT account, in the operating system’s key store;
• the log of the Codex core, which ArchX never sends;
• a list of recent error codes with their time, without any message text; it reaches TYTOS only in a report you send.

What goes to your AI provider

For each instruction, the Codex core in the app sends your instruction, the conversation so far, attachments, the content of files the agent reads, tool results and images to the AI provider you chose:

• with your ChatGPT account: directly from your device to OpenAI;
• with an API key: a compatibility layer on your own computer, reachable only at the local address 127.0.0.1, passes each request directly to the provider of the key: OpenAI, Anthropic, Google, Mistral (European location) or xAI. The key goes only to that provider.

The provider processes this data under your contract with it. TYTOS sees neither the content nor your keys and is neither controller nor processor for this processing.

What reaches TYTOS

Automatically, from the app:

• the licence check, when the app starts and every 24 hours: your licence email address, the installation ID, operating system, processor type, the versions of the app and of the core, and the device confirmation key;
• the start of the trial and the device confirmation: the same details, the version of the terms, the email address and 6-digit code you confirm and, if the link you used carried any, its referral code and campaign labels (Privacy Notice, section 4.12);
• update checks: the update channel, operating system and processor type.

Usage counts, off unless you opt in: ArchX can count, per day, how often the app started, how many specification files its tools created, whether the initial setup is complete and how often the core crashed, without any content. For practices in the United Kingdom, the service keeps no such counts, and the UK release of ArchX keeps counting switched off unless you switch it on after a clear explanation. Declining does not affect your licence.

Only when you choose:

• ‘Report a problem’: your text, the installation ID, the email address of your ChatGPT account if you are signed in with it, system and version details, your ChatGPT plan and remaining allowance, the licence status, whether a specification catalogue is connected, and up to 50 recent error codes; attachments only if you tick them: a screenshot of the app window and messages you select, at most 40;
• emails you send to support;
• what you enter when you buy licences and on ‘Manage seats’.

Never: project files, file names, paths, conversation text (unless you attach messages to a report), API keys or passwords. The interface between the app and the service does not accept such content.

The service runs at Vercel in Frankfurt (Germany), and its database is in Zurich (Switzerland). The Privacy Notice names every recipient and how long we keep each kind of data.

No tracking and no crash reporting

ArchX contains no analytics, tracking or crash-reporting component. The Codex core has its own features for usage statistics, metrics and feedback to OpenAI; the app switches them off each time it starts the core.

Other connections of the app

These sources see the IP address of your device; they receive no data from TYTOS:

• npm (registry.npmjs.org): new versions of the Codex core, only versions we have released;
• OpenAI’s sign-in pages, when you sign in with ChatGPT;
• GOV.UK: the list of bank holidays in England and Wales, when the agent writes a tender letter with dates;
• your practice’s own website, only after you have confirmed that the agent may read it;
• Hugging Face on Windows and Apple on a Mac, for the speech model and language packs, only for speech input;
• on Windows, spell-checking dictionaries from a Google server; according to Electron, no input is transmitted;
• plugin directories you add, and the browser tool, MCP servers, plugins and accounts you connect: they connect to the providers you choose, and you are responsible for these data flows.

Personal data of your clients and others: our recommendations

Drawings, letters and schedules can contain personal data, for example the names and addresses of clients and neighbours. For this data, your practice is the controller under UK GDPR, and your AI provider processes it under your contract with that provider.

• Choose a business plan. Use ChatGPT Business or Enterprise, or an API key from your practice’s account, and put the provider’s data processing terms in place. According to OpenAI, content from ChatGPT Business and Enterprise is not used for training by default, and OpenAI offers a data processing addendum for these plans and for its API (OpenAI: Enterprise privacy).
• With ChatGPT Plus or Pro, switch off the setting to improve the model in ChatGPT’s data controls; according to OpenAI, it also applies to tasks through Codex and therefore through ArchX (OpenAI: Data Controls). Plus and Pro are individual plans; we do not recommend them for documents with clients’ personal data.
• With an API key, check the provider’s terms. According to the providers, OpenAI, Anthropic, Google (paid tier) and xAI do not use API data for training by default; with Mistral, switch training off in your Mistral account.
• Transfers: your provider may process data outside the United Kingdom, for example in the USA. Check that your agreement with the provider covers such transfers.
• Tell your clients in your own privacy notice that you use AI services to prepare documents, and name the provider.
• Leave out what is not needed: do not give the agent special category data (Article 9 UK GDPR), such as health data, unless it is necessary and your provider’s terms allow it.
• Check every result: results are drafts. A named, suitably qualified person in your practice checks each result before it is used and takes professional responsibility for it (Licence Terms, section 15).