← All posts

What your dictation app sends besides your voice

Most privacy questions about dictation apps stop at the audio. Is my voice transcribed on my Mac or in the cloud? That is the right first question, but for a developer it is no longer the only one. Many AI dictation apps now read context to improve what they type: your clipboard, the text you selected, the text in the window you are typing into. That context goes to the language model that cleans up your words. If that model runs in the cloud, so does your context.

For someone who dictates prompts into Claude Code, Cursor or Codex all day, that matters. The last thing you copied might be a stack trace, a line from .env, or a customer’s email. This post shows the four things a dictation app can send, where two popular apps document them, and how to check your own.

Checked on 16 October 2026 against each vendor’s own documentation and privacy pages. Products change; read their current pages before you decide.

The four things a dictation app can send

Think of a dictation app as a short pipeline. Each step can run on your Mac or on someone else’s server.

  1. Audio. Your voice, turned into text by a speech model.
  2. Transcript. The raw words, which a language model may rewrite into clean prose or a structured prompt.
  3. Context. Extra text the app adds to help that language model: your clipboard, your selection, the active app’s name, the text in the field you are typing into.
  4. History and training data. Past transcripts, corrections and audio, kept on your Mac, synced, or shared to improve models.

An app can keep step 1 on your Mac and still send steps 2 and 3 to a cloud model. That is the case people miss. “Offline transcription” on a product page says nothing about where the clean up happens, or what goes with it.

What context capture looks like in practice

Superwhisper documents its context feature in detail, which makes it a good example. Its docs list three kinds of context:

  • Selected text, “captured exactly when you start recording”.
  • Clipboard, “captured either within three seconds before recording starts OR during dictation”.
  • Application context: “text from active input fields, names, and title from your active window”.

Its Super mode page adds that, with application context on, the app “includes system information, like the current date and time, your full name, and your computer name”. All of this goes to the language model you picked for that mode. The docs also show you how to see it: the History tab has a prompt field that “shows exactly what context was sent to the AI”.

Superwhisper is also clear that you can keep everything local. Its sensitive data page says that with a local voice model and a local language model, “no audio or text ever leaves your machine”. When you use its hosted cloud models, it says providers process data under zero data retention API terms, and that your audio and text “are not used to train any AI models”. So the question is not whether the app is safe. It is which models your modes use.

Wispr Flow works differently. Its data usage page says Flow “records the audio, which is then sent to our servers for processing”, so transcription itself is a cloud step. Its privacy policy says that with Context Awareness on, it may collect “limited, relevant content from the specific app in use (such as the text on the screen)”, and that you can opt out in settings. Its “Help improve Flow” setting decides whether audio, transcripts and edits are stored on its servers to train models.

None of this is hidden. It is all in the docs. But few people read them before they hold a hotkey over a terminal full of production logs.

Why this hits developers harder

A writer’s clipboard holds a paragraph. A developer’s clipboard holds:

DATABASE_URL=postgres://admin:...@prod-db.internal:5432/app
Traceback (most recent call last):
  File "/srv/billing/invoice.py", line 212, in charge_customer
ghp_************************************

You copy a failing log line, then dictate “fix this error”. A context feature working as designed may send that log line to a language model, which is the point of the feature. If the model is in the cloud, your log line went with it. The same goes for selected code in Cursor, or the text of the terminal window you are dictating into.

Your company may also have a rule about this. Many teams already decide which AI tools may see source code. A dictation app with cloud context is one more of those tools, even if it feels like a keyboard.

How to check your own dictation app

You can answer all four questions in about fifteen minutes.

  1. Open the settings for each mode you use. Note which speech model and which language model it runs. Anything labelled with a provider name, or “cloud”, sends text out.
  2. Look for context toggles. Search the settings for clipboard, selected text, application context, screen, or context awareness. Turn off the ones you do not need.
  3. Find the history view. If the app shows what it sent to the model, as Superwhisper does, dictate once with something harmless on your clipboard and see whether it appears.
  4. Check the training setting. Look for “improve the product” or “share data” toggles and decide on purpose.
  5. Watch the network. Turn off Wi-Fi and dictate in every mode, then watch traffic with Wi-Fi on. Our guide to checking if Mac dictation runs offline walks through Activity Monitor, nettop and lsof.

If a mode only works online, its text goes to a server. If it works offline, the text, and any context it reads, stays on your Mac.

Where Akova draws the line

We built Akova, on-device dictation for coding agents, so that the whole pipeline can stay on your Mac.

  • Audio is transcribed with WhisperKit, which runs Whisper models through Core ML on Apple Silicon.
  • Clean up and Prompt mode run locally: on Apple Intelligence’s on-device models on macOS 26, or a local Ollama model on macOS 14 and 15. Whatever text goes into Prompt mode stays on your Mac, because the model does.
  • History is off by default. Turn it on and up to 500 transcripts stay on your Mac, never uploaded.
  • Learned vocabulary (beta, off by default) reads your recent Claude Code session files on your Mac to pick up project words. It keeps only a short word list, locally, and nothing leaves your Mac.
  • Cloud transcription exists only as an opt-in, with your own provider key, sent straight from your Mac to that provider.

Lifetime also adds an option to redact personal data before it pastes. The full list of what the app and this website send is in our privacy policy. For the wider trade off between local and cloud speech, read on-device vs cloud dictation for Mac developers.

Akova needs macOS 14 or later on Apple Silicon (M1 and later). It is free to start, and Prompt mode is part of a one time Lifetime licence; see pricing for what each plan includes.

Your next step

Copy something harmless, like hello from the clipboard, then dictate a prompt in your usual app and mode. Open its history or watch its traffic. If your clipboard went somewhere you did not expect, change that setting before you copy your next secret.