Documentation

Quick start

Everything you need for the first ten minutes, plus the fixes for the few things that go wrong on a fresh machine. Written for 0.9.11.

Install

Download the build for your platform from the download page. macOS 12 or later on Apple silicon; Linux as AppImage or .deb. Windows is in development.

Linux

chmod +x Keypilot-Voice-0.9.11-x86_64.AppImage
./Keypilot-Voice-0.9.11-x86_64.AppImage

The macOS build is ad-hoc signed while notarization completes. If Gatekeeper blocks the first launch, right-click the app and choose Open once.

First run

01Enter your email address. We send a six-digit code — no password to remember.
02Paste the code. A new account is created automatically with 5,000 credits.
03Allow microphone access, then allow Accessibility (macOS) so text can be typed into other apps.
04Click into any text field, hold the dictation hotkey, and speak a sentence.

Speak at a normal pace. Streaming transcription starts before you finish the sentence, so the text is usually there the moment you release the key.

Hotkeys

Dictate⌘RightHold to talk, release to insert

Edit selection⌥RightSelect text first, then speak an instruction

Linux equivalentsCtrlAltRightRemappable in Settings

Hold-to-talk is the default. Switch to press-once in Settings if you prefer to tap and talk.

Credits & usage

• Speech recognition costs 1 credit per second of audio — about 60 per minute.
• AI rewriting costs 2 credits per 1,000 tokens.
• New accounts get 5,000 credits, refreshed monthly; purchased credits never expire.
• Every session is listed in the usage ledger with duration, model and cost.

Running low? The app tells you before a session starts rather than cutting you off mid-sentence. Top up from the pricing page.

Use your own models

The model hub can point speech recognition and rewriting at your own providers — for example a Volcano Engine, Alibaba Cloud or OpenAI-compatible endpoint. With your own keys, no credits are consumed and audio goes directly from your machine to that provider.

Keys are stored in the system keychain, never in the project folder or in logs.

Privacy

What leaves your machine

Audio, only while a dictation session is open, plus the hotword list that improves accuracy. Transcripts are returned to the app and written to a local database.

History, dictionary, settings and usage stats live in a local SQLite database. Export everything as JSON or delete it from the account page — the file is on your disk, not ours.

Troubleshooting

The transcript appears, but nothing is typed into the app: macOS needs Accessibility permission for text injection. Open System Settings → Privacy & Security → Accessibility and enable Keypilot Voice, then relaunch the app.
The capsule appears but stays empty: Check the input device in Settings and watch the level meter while you speak. A muted or wrong microphone is the usual cause.
“Insufficient credits” before a session: A session needs at least 2 credits to start. Wait for the monthly refresh, buy a pack, or switch to your own API key in the model hub.
Linux inserts nothing into the focused window: Some Wayland compositors block synthetic keystrokes. Install xdotool, or run the app under an X11 session.
The app will not open on macOS: Right-click → Open the first time. The build is ad-hoc signed until notarization completes; if it still refuses, remove the quarantine attribute with xattr -d com.apple.quarantine.

Hardware (optional)

Keypilot began as an open voice keyboard built around an ESP32-S3 with eight keys, a knob and a microphone. The desktop app can drive it over USB HID or Wi-Fi, so the push-to-talk key can live on your desk instead of your keyboard.

It is entirely optional — the software works with any microphone — and the hardware pages stay in the app for people who have a board.

Document version 0.9.11 · updated 2026-10-11

Report a problem