PromptStream

I designed and built PromptStream on my own: the Electron desktop app for Linux and macOS, the Next.js website and API, accounts with Supabase, and the hosting on my VPS.

• Wayland, the default on Fedora, doesn’t let one app type into another app’s window.
• It can’t rely on integrations with each app, so it pastes through the clipboard like a person would.
• Speech has to reach a transcription model, so dictation needs a network connection.

The desktop app is Electron with React and TypeScript, with separate macOS and Linux back ends for the global shortcut and for pasting.

While the shortcut is held, a small overlay window records the microphone and shows a live level meter. The app sends the audio to a Next.js API route with the user’s Supabase session token. That route transcribes it with Whisper on Groq, and a second route cleans up the text with an LLM, so the API keys stay on the server. The personal dictionary goes to both as a glossary, so names and jargon come out spelled right.

To paste, the app puts the text on the clipboard and sends the paste shortcut: through AppleScript on macOS, and through the RemoteDesktop portal over D-Bus on Linux. On macOS the previous clipboard comes back afterwards. History stays on the computer, capped at 500 entries.

On Linux, the usual ways to send a key press don’t work everywhere: XTEST through XWayland only reaches X11 apps, and tools like wtype or ydotool need a wlroots compositor or root access.

The XDG RemoteDesktop portal, which GNOME and KDE both implement, is the only route that reaches native Wayland windows. The cost is a permission prompt the first time on GNOME; the app asks the portal to remember the grant, so it only appears once.

• 24 Vitest unit tests cover the hotkey state machine, key combination tracking and the text helpers, and they pass.
• If cleanup or pasting fails, the transcribed text stays on the clipboard and in history, so nothing you said is lost.
• Requests time out (60 seconds for transcription, 25 for cleanup), and an expired session, a plan limit or a rate limit each get their own message.
• If the microphone only records silence, the app says so and, on macOS, asks for access again. If the chosen microphone disappears, it falls back to the default one.
• The LLM sometimes answered a dictated question instead of cleaning it up. Few-shot examples and a check that falls back to the raw text fixed it.
• Not done yet: the website and API have no tests.

PromptStream is released at promptstream.alabaganne.com for macOS 13+ and Linux (AppImage), with a free plan, and I update it from time to time. Windows isn’t supported yet.

• promptstream.alabaganne.com (opens in a new tab)
• Download page (opens in a new tab)

The code is in a private repository.