> ## Documentation Index
> Fetch the complete documentation index at: https://www.speakoflow.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Privacy

> What SpeakoFlow sends and what it keeps on your device: speech-to-text runs locally, there is no telemetry and no account, web search and personal memory are off by default, and API keys go to your OS credential store.

Four things are true no matter how you configure SpeakoFlow.

* **Your voice is transcribed on your device.** Speech-to-text always runs locally, on your GPU or CPU, and audio is never uploaded.
* **The assistant only contacts the provider you chose.** That can be the fully offline built-in llama.cpp engine, a local server like Ollama or LM Studio, or a cloud provider with your own key.
* **No telemetry and no account.** SpeakoFlow does not phone home.
* **Optional features stay off until you turn them on.** [Web search](/docs/assistant/web-search) and [personal memory](/docs/personalize/memory) are both off by default.

## What leaves your machine, and when

| What                                                                                                            | When                                                                         | Where it goes                            |
| --------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------- | ---------------------------------------- |
| Assistant, [AI cleanup](/docs/writing/ai-cleanup), [Generate with Flow](/docs/writing/generate-with-flow), spoken answers | Only when you use them, and only if you configured a remote provider         | The provider you chose, and nowhere else |
| [Web search](/docs/assistant/web-search)                                                                             | Off by default. When on, only when the assistant decides a search would help | Your chosen search backend               |
| Screen captures                                                                                                 | Only when you ask for them                                                   | The provider you chose                   |
| Model downloads                                                                                                 | When you download a transcription or local assistant model                   | The model host                           |
| Update checks                                                                                                   | **On by default.** Toggle in *Settings → General*                            | This project's GitHub releases page      |

Pick a local provider for the first three rows and the app never opens a connection at all.

## What is stored on this device

<AccordionGroup>
  <Accordion title="Screen captures">
    Screen vision captures nothing unless you ask it to, and then only the monitor under your mouse cursor.

    The full-resolution frame is sent once to the model provider you chose and is not kept. What persists locally is a small thumbnail attached to the message, so you can see later what the assistant was looking at. Details in [Screen vision](/docs/assistant/screen-vision).
  </Accordion>

  <Accordion title="Personal memory">
    Off by default. When you turn it on, everything is stored on this device in your settings file. *Settings → Assistant → Memory* lets you read, edit, export, or erase every note.

    Memory is advisory only, so injected memory never overrides what you asked for in the current message. It actively refuses to keep secrets, personal identifiers, or instruction-shaped text at capture, consolidation, and injection time.

    An Incognito toggle skips both using and learning memory for a single conversation.
  </Accordion>

  <Accordion title="History and recordings">
    *Settings → History* keeps your recent dictations, Flow generations, and assistant chats on this device. Two settings control how much is kept:

    * **Recording limit**. How many recordings to keep. Default 20.
    * **Auto-delete recordings**. `Keep a set number` (the default, which means the recording limit above does the trimming), `Never`, `After 3 days`, `After 2 weeks`, or `After 3 months`.

    Both apply only to unstarred dictation and Flow recordings. Starred recordings are always kept, and assistant chats are never counted or auto-deleted.

    Changes apply immediately, so lowering the limit deletes the excess right away.
  </Accordion>

  <Accordion title="API keys">
    Provider API keys go into the OS credential store, not into a plain-text settings file: Windows Credential Manager, the macOS Keychain, or the Linux Secret Service, under the service name `com.abhishekbarali.speakoflow`.

    They are read back into memory once at load and cached there, so the keychain is never touched in the hot path.

    <Note>
      If the platform store is not available, a headless Linux box with no Secret Service for example, the app keeps the key in its settings file rather than losing it. On such a machine the key is on disk in plain text, so it is worth knowing.
    </Note>
  </Accordion>

  <Accordion title="Keyboard access">
    Global shortcuts have to observe your keyboard to know when you pressed one. Nothing logs or stores what you type.

    On Linux the handy-keys engine reads `/dev/input/event*` directly, which is why it needs your user in the `input` group. The Tauri engine, the default on Linux, uses the compositor's global-shortcut API and reads nothing.
  </Accordion>
</AccordionGroup>

See [How SpeakoFlow works](/docs/start/how-it-works) for the full data flow, and [Providers & keys](/docs/assistant/providers) for what gets sent where when you use a cloud model.