Ambient, local-first AI co-pilot for live conversations on your Mac.
Listens, transcribes on-device, and lets you ask an AI about what's happening — without alt-tabbing.
Whisper Pilot listens to anything your Mac can hear — meetings, podcasts, tutorials, your own voice — transcribes it on-device, and streams answers from your favorite LLM into a translucent floating overlay. Bring your own key. No backend. No telemetry. No signup.
Built for: live meetings (Zoom, Meet, Teams, Slack, Discord), tutorials and lectures, pair programming with screen context, live translation, and anything else you might want to ask a question about right now.
Alpha note: releases are not yet signed with an Apple Developer ID, so macOS Gatekeeper will block the first launch. See Allow the app through Gatekeeper below for the one-time unblock — it takes about ten seconds. Prefer to build from source? See
docs/CONTRIBUTING.md.
brew install --cask vertocode/whisper-pilot/whisper-pilotUpdate with brew upgrade --cask whisper-pilot. Uninstall with brew uninstall --cask whisper-pilot. Pass --zap to also clear preferences (session transcripts are deliberately kept).
Grab the latest .dmg from the Releases page and drag WhisperPilot.app into /Applications.
Because Whisper Pilot isn't signed with an Apple Developer ID yet, the first launch (whether installed via Homebrew or the .dmg) shows a dialog like "WhisperPilot can't be opened because Apple cannot check it for malicious software." This is expected.
To allow it:
- Open System Settings → Privacy & Security.
- Scroll down to the Security section — you'll see a line that says "WhisperPilot was blocked to protect your Mac."
- Click Open Anyway next to that line. macOS asks you to confirm with your password / Touch ID.
- Launch Whisper Pilot again. This time a smaller dialog appears with an Open button — click it once. From then on, the app launches normally with no prompts.
If you prefer the Terminal: xattr -dr com.apple.quarantine /Applications/WhisperPilot.app removes the quarantine attribute and skips the dialog entirely.
- Launch — the Sessions window opens. Click Start new to enter the overlay.
- Open Settings from the overlay's
…menu → AI Provider tab → paste your Gemini API key. Stored in Keychain. - Click ▶ in the overlay. macOS will prompt for Microphone permission (so your own voice can be transcribed) and, on Macs where the Core Audio Process Tap can't capture system audio, Screen Recording as well (used only for the system-audio capture path — no video is recorded). Grant whichever it asks for.
macOS's privacy system (TCC) binds permission grants like Microphone and Screen Recording to the code signature of the binary that asked for them, not just the bundle ID. Whisper Pilot is currently distributed unsigned (no Apple Developer ID), so every release produces a different code-signing identity (none, really) — macOS sees each new build as a different app and asks for permission again.
There's no script-level workaround: macOS deliberately doesn't let installers or sudo grant TCC permissions on the user's behalf. The fix is to sign the app with a stable Apple Developer ID, after which TCC keeps the grant across rebuilds. bin/release already supports this when WP_DEVELOPER_ID is set in the environment — adopting it just requires an Apple Developer Program membership ($99/year).
If your Mac can't see system audio (typical on some Mac mini / USB / Bluetooth output configurations where the default Core Audio Process Tap silently delivers no frames), Whisper Pilot now detects it during the session and switches to the ScreenCaptureKit capture path automatically — silently when Screen Recording permission is already granted, or via a one-click prompt in the transcript pane otherwise. The choice is remembered for future sessions.
If Whisper Pilot vanishes without warning ("the app closed by itself"), it's almost always because of a hard crash. Because the app runs as an accessory (no Dock icon), macOS doesn't always show a dialog. Two places to look:
~/Library/Application Support/com.whisperpilot.app/runtime.log— Whisper Pilot's own rolling log, written line-by-line. The tail is what was happening immediately before the crash. The app surfaces the last ~20 lines automatically on the next launch via the in-app log pane.~/Library/Logs/DiagnosticReports/WhisperPilot-*.crash(or.ips) — macOS's own crash report with a full stack trace. Open in Console.app (Action → Reveal in Finder works backwards too).
Please attach both when reporting a bug about unexpected closes.
Hit a problem during install or use? Please open an issue — include your macOS version, what you ran, and any messages from the overlay's log panel. Pull requests with a fix are even better.
- macOS 14 (Sonoma) or later
- Apple Silicon recommended (Intel works; transcription is slower)
- A Gemini API key from aistudio.google.com
The overlay is always available once you've picked a session.
- Detected questions. When someone asks you a question in a meeting, the trigger engine notices and streams a suggested answer into the AI lane automatically.
- Composer. Type any question in the box at the bottom. ⌘⏎ to send. Tick See my screen to also include a screenshot.
- Auto-send. Optionally set an interval (30s–5m) and the assistant proactively summarizes the recent conversation on every tick. Off by default.
- Pause AI. Sparkles button in the header. Listening + transcribing keep running; only the AI is silenced.
- Audio never leaves your device. Capture and transcription are entirely local. English sessions use the Parakeet Unified 0.6B CoreML model (via FluidAudio) running on the Neural Engine — the same accuracy class as Meet/Teams server captions; it downloads once (~600 MB) on first use. Other locales use Apple's on-device speech engines.
- The LLM is the only thing that talks to the network, and only when you trigger it (auto-send tick, detected question, or composer submit). No background polling.
- Your API key lives in the macOS Keychain. Never written to disk in plaintext.
- Screenshots are sent only when you tick See my screen — never cached, never persisted.
docs/ARCHITECTURE.md |
Module-by-module breakdown, data flow, protocols |
docs/CONFIGURATION.md |
Every setting and what it controls |
docs/SESSIONS.md |
On-disk session format and resume semantics |
docs/CONTRIBUTING.md |
Build from source, conventions, areas that need help |
docs/RELEASE.md |
DMG signing, notarization, Homebrew tap workflow |
docs/ROADMAP.md |
What's next: WhisperKit, Ollama, diarization, modes, RAG |
The codebase is small, protocol-first, and intentionally non-magical — most contributions fit inside a single module. Start with docs/ARCHITECTURE.md, then docs/CONTRIBUTING.md. High-value areas right now: multilingual Parakeet/Nemotron engines, Ollama provider, snapshot tests for the trigger heuristics, real app icon.
MIT — see LICENSE.