> ## 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.

# How SpeakoFlow works

> The mental model behind SpeakoFlow: one hotkey layer over the app you are already in, a dictation pipeline that runs entirely on your machine, and an assistant you can run offline, on your own server, or in the cloud.

SpeakoFlow sits in your system tray and does nothing until you press a shortcut. You never switch to it to dictate. It is a layer over whatever app you are already in.

Three jobs live under that layer.

* **[Dictation](/docs/dictation/basics)**. Local speech-to-text, typed into the active app.
* **[Generate with Flow](/docs/writing/generate-with-flow)**. Start a dictation with an activation phrase and the AI writes the finished result.
* **[The assistant](/docs/assistant/panel)**. A floating panel you summon to ask questions by voice or text.

## The dictation pipeline

<Steps>
  <Step title="Your microphone" icon="microphone">
    Audio is captured from the input device you picked in [General settings](/docs/settings/general).
  </Step>

  <Step title="Voice-activity detection" icon="filter">
    Silero VAD filters out silence, so pauses while you think don't become gaps in your text.
  </Step>

  <Step title="Speech-to-text, on your hardware" icon="microchip">
    whisper.cpp or Parakeet runs the model on your GPU, or your CPU if no GPU is available. See [Acceleration & GPU](/docs/models/acceleration).
  </Step>

  <Step title="Text into the active app" icon="clipboard">
    Inserted via the clipboard or typed directly. See [Output & pasting](/docs/dictation/output).
  </Step>
</Steps>

Nothing in that chain touches the network.

Two things can add a network hop after it, and both are opt-in: [AI cleanup](/docs/writing/ai-cleanup) and [Generate with Flow](/docs/writing/generate-with-flow). They hand the text to whichever model provider you configured, which may be the fully local built-in engine.

## What the assistant does with a question

An assistant turn is not a single request. The model gets a small set of tools, web search and the current date and time, and decides for itself whether it needs them. The loop is bounded, so a turn can't run away. With web search off, the tool isn't offered at all and the assistant answers from what it knows.

[Screen vision](/docs/assistant/screen-vision) works the same way. Nothing is captured unless you ask for it.

## Where things run

Speech-to-text is always local. The assistant is where you choose your comfort level, in *Settings → Assistant*.

* **Built-in**. A small local model, fully on your machine. No key, no network.
* **Local server**. Ollama or LM Studio on your own machine.
* **Cloud**. Your own API key for any OpenAI-compatible provider.

See [Models and providers](/docs/assistant/providers) to set this up, and [Privacy](/docs/reference/privacy) for exactly what is stored and where.