Docs

Privacy, AI and Integrations

Network Transparency

StratoSort Stack's headline claim is that your files and their contents never leave your machine. This document makes that claim verifiable: it enumerates every network connection the shipped app can make, the exact conditions that trigger each one, and a test you can run yourself. Every row is backed by a source reference so the claim can be re-audited against the code.

The short version: analysis, organization, search, chat, and transcription run entirely on-device. The app touches the network only to (1) download AI models (the default text and embedding models download automatically at first launch unless you pick "Continue with limited AI"), (2) check GitHub Releases for app updates (on by default in installed builds; switch it off in Settings), and (3) call cloud AI providers if — and only if — you configure them with your own API keys. There is no telemetry, no crash upload, no account, no phone-home.

Complete network surface

Connection When it happens Source
https://huggingface.co (model files) Only when you download or update an AI model. Host allowlist: huggingface.co, hf.co; HTTPS only; no credentials sent. src/shared/modelRegistry.js, src/main/services/ModelDownloadManager.js
App updates (STRATOSORT_ENABLE_AUTO_UPDATES) Packaged builds only: at startup and every 4 hours while Settings → Application → Background Update Checks is on (the default), plus when you click "Check for updates". Fetches release metadata from the packaged GitHub Releases feed, or from STRATOSORT_UPDATE_FEED_URL if set (HTTPS, allowlisted host required). STRATOSORT_ENABLE_AUTO_UPDATES=false disables it entirely. No file or document data is sent. src/main/core/autoUpdater.js
Cloud AI providers (Anthropic, or any OpenAI-compatible endpoint) Only if you enter an API key and select that provider in Settings. Off by default; the default provider is the local model. There are two provider implementations: Anthropic (api.anthropic.com) and OpenAI-compatible, which reaches whatever base URL you enter — OpenAI by default, or Groq, DeepSeek, a self-hosted server and so on. Any HTTPS host is accepted; plain http:// only for localhost, 127.0.0.1 and [::1]. src/main/services/ai/providers/, src/main/services/ai/providerBaseUrl.js
Local inference servers (Ollama localhost:11434, LM Studio localhost:1234) Only if you set one as the OpenAI-compatible base URL. One option among the endpoints above; when the URL is loopback, the traffic stays on your machine. AI provider settings
Built-in MCP server (http://127.0.0.1:<port>/mcp) Only if you enable it in Settings. Binds to 127.0.0.1 exclusively — never reachable from the network. The AI tool you connect to it (Claude Desktop, Cursor, Zed) receives search results, summaries, passages and answers from your library and sends them to its own model, which may be in the cloud; that traffic is the client's, not StratoSort's. src/main/transports/mcp/transport.js, docs/MCP_SERVER.md
https://huggingface.co (core models at first launch) At first launch the default text and embedding models download automatically, unless you pick "Continue with limited AI" on the welcome screen. Later launches re-download a core model that is missing from disk, under the same opt-out. The vision model is never fetched here — only from the setup wizard. Same host allowlist and checksum rules as a manual download. src/main/core/backgroundSetup.js
https://huggingface.co (Whisper speech model) Only when you click Install model on the banner that appears after an audio or video transcription finds ggml-base.en.bin (about 150 MB) missing. Downloaded from the ggerganov/whisper.cpp repository over HTTPS with the same host allowlist; the expected SHA-256 is read from that repository's Git LFS pointer at download time rather than pinned in the app. src/main/services/WhisperModelInstaller.js, scripts/setup-whisper-model.js
https://github.com (vision runtime binary) Only if the bundled llama-server binary is missing (a development checkout, or a damaged install). Host checked on every redirect; SHA-256 pinned. src/main/services/VisionService.js
https://cdn.jsdelivr.net (OCR language data) Only if the bundled eng.traineddata is missing (a development checkout, or a damaged install). Packaged builds ship it, staged at build time from a pinned package version and verified against a pinned SHA-256; the runtime fallback download is not hash-checked. src/main/utils/tesseractJsPaths.js, scripts/setup-tessdata-runtime.js

What is deliberately absent:

The airplane-mode test

You can verify the offline claim end-to-end in about two minutes:

  1. Install StratoSort and download a model profile while online (the one network step). If you plan to transcribe audio, also install the Whisper model while online.
  2. Disconnect completely — enable airplane mode, or unplug Ethernet and disable Wi-Fi.
  3. Use the app: analyze documents and photos, transcribe an audio file, organize into smart folders, run semantic searches, chat with your documents.
  4. Everything works. No feature degrades, no errors about connectivity, nothing queues waiting for a connection.

For a stronger check, run a local firewall monitor (e.g., Windows Defender Firewall logging, Little Snitch on macOS, OpenSnitch on Linux) during normal indexing and search: with no cloud provider configured and no update feed set, the app makes no outbound requests.

Model download integrity

Every model in the registry ships with a pinned SHA-256 checksum, and the downloader refuses to install a file that does not match:

To audit the pinned hashes yourself, read the registry file — model names, exact URLs, sizes, and checksums are all declared in one place — and compare against the SHA-256 values published on the corresponding HuggingFace repository pages.

Logs and crash dumps

Structured logs (JSONL) and crash dumps stay under the app's user-data directory. Log content is sanitized (secrets such as tokens and API keys are redacted by src/shared/logger.js), and nothing is transmitted anywhere. Deleting the user-data folder removes all of it.

Edit this page on GitHub