Privacy Policy

Last updated: July 2026

Introduction

Screph ("we", "us" or "our") respects your privacy and is committed to protecting your personal data. This Privacy Policy explains how we collect, use and protect your information when using our website screph.com and Screph software.

What Data We Collect

We may collect the following types of information:

  • Website request log: for successful GET/POST requests to public pages, the server stores the path, method, timestamp, IP address, User-Agent, referrer, bot flag and, for a signed-in user, an account reference. Static/media, administrative, API, account-login and OAuth routes are excluded from this general log.
  • Account data: If you register on the site - username, email and other data you provide.
  • Download events: when a tracking row is created successfully: the served file name and size, transfer status, bytes sent, completion time, IP, User-Agent and referrer. Django streaming updates progress as the iterator is consumed. With the production Nginx X-Accel-Redirect handoff, the site records only the start; without separate reconciliation the row can remain started with zero bytes and no completion time. An emergency raw-file fallback after response preparation fails may serve the file without a DownloadEvent. The implementation does not measure page-view duration.
  • IP metadata: the server may attach previously stored country, region, city and organization data for the same IP. The website does not send the IP to an external geolocation service while processing the request.
  • Feedback data: Text and explicitly attached images submitted through support chat, plus information you provide when participating in the community. While the desktop chat is visible it also fetches the thread about every 20 seconds, sending the account token or guest key; hiding the widget stops polling.
  • Online operation data: scoped device authorization; a periodic authenticated profile request when a token is stored; WAV audio for a selected server-speech operation, its duration and usage ledger; an explicitly submitted diagnostics ZIP and metadata; context for a selected LLM/speech provider; or a model-list request or synthetic silent probe.wav when OpenAI/LiteLLM Speech settings are opened.

How We Use Data

We use collected data to:

  • Provide and improve our services.
  • Analyze usage and optimize site performance.
  • Contact you regarding support and updates.
  • Ensure security and prevent abuse.

Cookies

The site uses functional Django cookies for sessions, CSRF protection and language. localStorage may contain the selected language, display name and API authToken; the password is not stored there. Clearing browser storage ends the local API session and resets these preferences. The current public templates do not load a third-party analytics script: the server-side request log described above does not rely on an analytics cookie.

Data Sharing with Third Parties

We do not sell or share your personal data with third parties, except in the following cases:

  • When required by law.
  • To protect our rights and user safety.
  • The IP lookup provider receives the public IP when server-side log enrichment is performed.
  • The browser loads Bootstrap CSS/JS from cdn.jsdelivr.net; the CDN receives standard HTTP request metadata.
  • Pages with embedded demonstrations load the YouTube player when opened; YouTube/Google receives normal browser network metadata and may apply its own cookies and policies.
  • Cloud LLM/speech providers, model hosts (including Alpha Cephei for the Vosk catalogue and archives) and stream URLs selected by the user receive data for the corresponding request under their own policies.

Data Security

We apply reasonable technical and organizational measures to protect your data. However, data transmission over the internet cannot be completely secure, and we cannot guarantee absolute protection.

Your Rights

In accordance with applicable law, you may have the right to:

  • Access your personal data.
  • Correct inaccurate data.
  • Delete your data.
  • Restrict data processing.

For a request containing personal data, use the private support chat in your profile. Do not publish tokens, diagnostics or other sensitive data in a public GitHub issue or Telegram chat.

Local processing and external services

Markup, projects and local CV processing run on the user's computer. When cloud speech recognition, an LLM or another external service is explicitly selected, the chosen data is sent to that provider under its terms. Review the provider and outgoing context before sending.

Separately from user content, an installed build may check Updates.xml on startup by default, no more than once every 24 hours; its User-Agent includes Screph/<version> update-check, and the check can be disabled in settings. About 1.2 seconds after startup, the app may also fetch six HTML pages for its local embedded-help cache; the current guard permits another attempt after one hour and there is no separate toggle. Runtime-pack and CV-model downloads are separate actions. The Vosk model catalogue, however, is requested from alphacephei.com when its selector is first opened and cached for five minutes; the selected archive is downloaded only on button press. Projects, frames and voice recordings are not sent to these metadata/download hosts. A stream URL, however, connects the device to that stream endpoint.

The production configuration includes a SCREPH_TELEMETRY flag, but the current runtime initializes no Sentry or other telemetry exporter and no consumer of that flag was found. The flag alone therefore does not imply automatic desktop telemetry upload.

The Account widget schedules its first refresh about 100 ms after its own initialization. With a stored desktop access token, the app then requests the Screph profile/balance every 30 seconds; failures introduce 60, 120 and 300-second backoffs and a 401 may trigger OAuth token refresh. Opening OpenAI Speech settings with an available key performs an authenticated model-list GET. For a configured non-OpenAI LiteLLM Speech path, Screph instead sends a generated 0.1-second 16 kHz mono silent probe.wav. It is not a microphone recording or project content, but the provider may record the request in logs or usage.

Diagnostics and crash reports

Crash sessions and diagnostics bundles are created locally. Submitting a bundle to the server requires an authenticated account and an explicit user action or consent in the startup prompt. The Logging page shows a path-and-size manifest, not file contents; the startup prompt does not show that manifest and submits immediately after Yes. Projects and autosaves are not attached automatically, although logs may contain system paths, project names and workspace details. See the diagnostics guide.

Account services process data for selected features and maintain a stored session through periodic profile refresh; a visible support widget separately polls its thread. See the account and online services guide for the exact list and current limitations.

Changes to This Policy

We may update this Privacy Policy from time to time. The current version is always available on this page with the date of the last update.

Retention and deletion

The current implementation does not define one automatic deletion period for request/download logs, support threads and diagnostics reports. These records remain until administrative deletion or an applicable user request is processed. Local retention is component-specific: Assistant action history uses its saved retention when the runtime is created, while session-directory cleanup currently uses an effective 180-day window when the workspace opens and keeps active and unresolved sessions. That cleanup does not remove the project-local .screph_ai_assistant mirror, and the current UI toggle does not stop all generic workspace mirror writes. Projects, models, most caches and artifacts remain until component-specific or manual cleanup. The local append-only ScrephData/llm/usage_events.jsonl stores profile/model/token/cost/latency metadata for successful shared-runtime LLM responses without prompt or response text and without automatic retention; crash reports keep the newest 10 by default. Account tokens and API keys use the system credential store, while ordinary paths, endpoints and preferences use unencrypted settings JSON. External providers define their own retention. See the local-data guide for the full map.