Privacy Policy

Last updated: 18 November 2026

Summary

railreader2 does not collect, transmit, or store any personal data. No analytics, no telemetry, no accounts, no crash reports. The application runs entirely on your device by default.

There are two exceptions, both off by default and both under your control: the optional Vision Language Model (VLM) integration, which sends data only to an endpoint you explicitly configure (see Vision Language Model integration), and AI tools you connect to your library, which then read your notes and documents’ text and pass what they read to their AI provider (see AI tools).

Data stored locally

railreader2 stores the following data on your computer only, in the platform-standard application data directory:

None of this data leaves your computer unless you connect an AI tool (see AI tools). Help → Clean Up Temp Files removes caches, temporary files and old logs; the library index can be cleared in Settings > Library. To delete everything the application has stored, remove its data directory:

Network access

By default, railreader2 makes no network requests. PDF rendering, layout analysis (via the bundled ONNX model), and all other features run entirely on your device. The application does not send crash reports or usage statistics.

Network activity happens only in these cases, each of which you start or agree to:

The MCP server that AI tools connect to (below) makes no network requests itself: it talks only to the program that started it, over that program’s input and output.

Vision Language Model integration

railreader2 provides an optional Copy as LaTeX feature (keyboard shortcut Ctrl+L) and a corresponding CLI subcommand (railreader2-cli vlm) that send cropped images of detected layout blocks (equations, tables, figures) to an OpenAI-compatible Vision Language Model API for transcription. This feature is disabled by default and requires explicit user action to enable:

  1. Open Settings > VLM and enter an endpoint URL, model name, and (optionally) an API key.
  2. Press Ctrl+L, use the context menu, or invoke the CLI subcommand to send a block to the configured endpoint.

You choose the endpoint. It may be a fully local server (such as Ollama or vLLM running on your own machine), in which case no data leaves your device, or a cloud API (such as OpenAI, Google Gemini, or any other OpenAI-compatible provider), in which case the cropped block image and prompt text are transmitted to that provider.

railreader2 is not affiliated with, and is not responsible for, any third-party AI service you choose to use. Before enabling this feature with a cloud endpoint, you should:

If you want the benefits of the feature without sending any data off your machine, use a locally hosted endpoint (Ollama, vLLM, LM Studio, or similar). See the VLM setup guide for options.

To disable the feature entirely, leave the VLM settings blank (or clear them via Settings > VLM > Reset). With no endpoint configured, railreader2 makes no VLM network requests and the keyboard shortcut becomes a no-op.

AI tools (MCP)

railreader2 includes an MCP server that lets AI tools such as Claude Code, Claude Desktop and Codex read your library. It is off by default. Until you tick Let AI tools read my library in Settings > Library > AI Tools (or run railreader2-mcp enable), it refuses every request, even from a tool that has been set up.

What a connected tool can read: your notes and comments (and others’ comments written into your PDFs), with their replies, authors and colour categories; your named bookmarks; the page text of the documents you have opened; and each document’s title, authors, file name and page count.

What it can never read: file paths or folder names, the PDF files themselves or any images of them, anything from password-protected documents, and anything from documents or folders you have hidden (Hide from AI Tools in the Library window or on the start screen’s recent documents, Hide a Folder… in Settings). The server only reads; it cannot change your documents, notes or settings.

Where it goes: what a tool reads becomes part of your conversation with it, and is sent to that tool’s AI provider (for example, Anthropic for Claude and OpenAI for Codex). Their privacy policies and terms apply, and railreader2 is not responsible for them. As with the VLM integration, consider the sensitivity of your documents, and whether your institution permits sending their content to an AI service, before connecting a tool.

To stop it, untick Let AI tools read my library: requests are refused from then on, including from tools that are already running. Unticking a tool’s switch removes railreader2’s entry from that tool’s settings.

Third-party services

railreader2 does not integrate with any advertising networks or analytics platforms. Its only third-party integrations are the user-configurable VLM endpoint and the AI tools you choose to connect, both described above and both off by default.

Children's privacy

railreader2 does not collect any data from any user, including children.

Changes to this policy

If this policy changes, the updated version will be posted at this URL. Since railreader2 collects no data, substantive changes are unlikely.

Contact

For questions about this privacy policy, open an issue on GitHub.