Skip to content

Frequently Asked Questions

Last Updated: 2026-02-13

Common questions about VaulType — privacy, compatibility, features, and troubleshooting.


VaulType is a privacy-first, macOS-native speech-to-text application that lets you dictate text into any app where your cursor is active. It uses local AI models (whisper.cpp for speech recognition and llama.cpp for text post-processing) — nothing ever leaves your Mac.

The name captures the core philosophy:

  • Hush — Your voice stays private, never leaving your device
  • Type — Text appears at your cursor, as if you typed it

Yes. VaulType is open source under the GPL-3.0 license. It’s free to download, use, and modify.

VaulType supports 90+ languages through Whisper’s multilingual models (small, medium, large-v3), including English, Turkish, Spanish, French, German, Chinese, Japanese, Arabic, and many more. English-only models are also available for faster performance. Language can be auto-detected or manually selected.

Yes, 100%. VaulType is designed to work completely offline. All speech recognition and text processing happen locally on your Mac. The only network requests are:

  • Model downloads (one-time, optional)
  • Auto-update checks via Sparkle (can be disabled)

No, absolutely not. VaulType processes all audio locally using whisper.cpp. Your voice data never leaves your Mac. There is no cloud component, no API calls, and no telemetry.

By default, no. Text is transcribed, optionally processed by the LLM, and injected at your cursor — then discarded from memory. If you enable Dictation History (opt-in), transcriptions are stored locally in SwiftData on your Mac.

No. VaulType has zero telemetry, zero analytics, and zero data collection by default. If you opt in to crash reporting (Sentry), only crash logs are sent — never audio, text, or usage data.

Yes. Since everything is local, VaulType is suitable for sensitive work including:

  • Legal documents
  • Medical notes
  • Confidential business communications
  • Personal journaling

🔒 VaulType cannot inject text into password fields — this is a deliberate security measure.

What permissions does VaulType need and why?

Section titled “What permissions does VaulType need and why?”
PermissionWhy
AccessibilityTo inject text into the active app via CGEvent
MicrophoneTo capture your voice for speech recognition
Automation (optional)For voice commands that control other apps via AppleScript

See PERMISSIONS.md for a detailed breakdown.


VaulType runs on any Mac with macOS 14.0 (Sonoma) or later:

  • Apple Silicon (M1/M2/M3/M4) — Recommended. Metal GPU acceleration makes speech recognition 3-10x faster.
  • Intel Macs — Supported with CPU-only inference. Use smaller models (tiny, base) for acceptable speed.
ConfigurationMinimum RAM
Whisper tiny (Raw mode)8 GB
Whisper small (Raw mode)8 GB
Whisper small + LLM 3B16 GB
Whisper medium + LLM 3B16 GB
Whisper large-v3 + LLM32 GB

💡 Recommendation: 16 GB for the best experience with both STT and LLM features.

ModelSize
Whisper tiny~75 MB
Whisper base~142 MB
Whisper small~466 MB
Whisper medium~1.5 GB
Whisper large-v3~3.0 GB
Qwen2.5-3B Q4~2.0 GB
Phi-3.5-mini Q4~2.5 GB

You only need one Whisper model and optionally one LLM. Typical installation: 500 MB to 3 GB.

Yes. VaulType is tested on macOS 14 (Sonoma) and macOS 15 (Sequoia).


Yes! This is one of VaulType’s best use cases. Use Prompt Mode to dictate into your terminal, and VaulType will automatically format your messy speech into clear, well-structured prompts for Claude Code.

Example: You say “I want to refactor the authentication module to use JWT tokens instead of sessions and add proper error handling” and VaulType formats it into a clean, structured prompt before injecting it into the terminal.

Yes. VaulType injects text into any macOS app, including Terminal.app, iTerm2, Warp, Alacritty, and Kitty. Use the Clipboard injection method for the best terminal compatibility.

ModeWhat It Does
RawNo LLM processing — inject speech exactly as recognized
CleanFix spelling, grammar, punctuation, capitalization
StructureFormat as bullet lists, numbered steps, or tables
PromptReformat into clear, well-structured AI prompts
CodeFormat as code comments, commit messages, docstrings
CustomYour own prompt templates for any formatting need

Yes. VaulType has a full prompt template system. You can:

  • Write custom system and user prompts
  • Use variables like {text}, {language}, {app_name}, {context}
  • Import and export templates
  • Assign modes to specific hotkeys or apps

Yes, two modes are available:

  • Push-to-talk — Hold the hotkey to record, release to process
  • Toggle mode — Press to start recording, press again to stop

Can I use different modes for different apps?

Section titled “Can I use different modes for different apps?”

Yes. App-Aware Context (Phase 3) lets you configure default modes per application. For example:

  • Terminal → Prompt Mode
  • Mail → Clean Mode
  • Notes → Structure Mode
  • VS Code → Code Mode

Yes (Phase 4). You can:

  • Launch and switch between apps: “Open Safari”, “Switch to Terminal”
  • Manage windows: “Move window to left half”
  • Control system: “Volume up”, “Mute”, “Do Not Disturb on”
  • Chain commands: “Open Terminal and run npm start”
  • Create custom aliases: “Deploy” → run your deploy script

ModelSpeedAccuracyBest For
tinyFastestGood enoughQuick notes, Intel Macs
baseFastGoodGeneral use on Intel
smallBalancedVery goodRecommended for most users
mediumSlowerExcellentAccuracy-critical, long dictation
large-v3SlowestBestMaximum accuracy, M-series Pro/Max

💡 Start with small — it’s the sweet spot of speed and accuracy for Apple Silicon.

Which LLM should I use for post-processing?

Section titled “Which LLM should I use for post-processing?”
ModelSizeSpeedQuality
Qwen2.5-3B Q42 GBFastRecommended
Phi-3.5-mini Q42.5 GBFastGreat for English
Llama-3.2-3B Q42 GBFastGood multilingual

On Apple Silicon with Metal GPU acceleration:

Model5s Audio30s Audio
tiny~0.3s~1.5s
base~0.5s~2.5s
small~1.0s~5s
medium~3s~15s
large-v3~8s~40s

Measured on M1 MacBook Pro. M2/M3 are ~20-40% faster.

Can I use Ollama instead of built-in llama.cpp?

Section titled “Can I use Ollama instead of built-in llama.cpp?”

Yes. VaulType supports Ollama as an alternative LLM backend via its localhost REST API (port 11434). This is useful if you already have Ollama installed with models downloaded.

Configure in Settings > Advanced > LLM Backend > Ollama.


Most common causes:

  1. Accessibility permission not granted — Check System Settings > Privacy & Security > Accessibility
  2. Cursor not in a text field — Click into an editable area first
  3. Wrong injection method — Try switching between CGEvent and Clipboard in Settings > Advanced

See TROUBLESHOOTING.md for detailed solutions.

You need to download a Whisper model:

  1. Open VaulType Settings > Models
  2. Click Download next to a model (recommended: small)
  3. Wait for the download to complete

Try these in order:

  1. Use a smaller model (tiny or base)
  2. Ensure Metal GPU is enabled (Settings > Advanced)
  3. Close GPU-intensive apps
  4. Use push-to-talk for shorter recordings
  1. Check if VaulType is running: look in Activity Monitor
  2. Relaunch from /Applications/VaulType.app
  3. Check System Settings > Control Center > Menu Bar Only to ensure it’s not hidden
  1. Verify microphone permission is granted
  2. Select the correct input device in Settings > Audio
  3. Check that the microphone works in Voice Memos

How is VaulType different from Apple’s built-in Dictation?

Section titled “How is VaulType different from Apple’s built-in Dictation?”
FeatureVaulTypeApple Dictation
Privacy100% localPartial (on-device + cloud)
LLM post-processingYes (6 modes)No
Voice commandsYes (app launch, window mgmt)Limited (system commands only)
Custom modesYesNo
Open sourceYes (GPL-3.0)No
Offline supportFullPartial (basic offline mode)
Text injectionAll appsAll apps
App-aware contextYesNo
CostFreeFree with macOS

How is VaulType different from Superwhisper / VoiceInk?

Section titled “How is VaulType different from Superwhisper / VoiceInk?”
FeatureVaulTypeSuperwhisperVoiceInk
PricingFree (GPL-3.0)$8/month$30 one-time
Dual AI pipeline (STT + LLM)Yes (both local)Partial (local + cloud)No
Voice commandsYesNoNo
Processing modes64+Basic
Zero-networkYesPartialPartial
Open sourceYesNoNo
Developer-focused modesYes (Code, Prompt)NoNo
App-aware contextYesYesNo

How is VaulType different from MacWhisper?

Section titled “How is VaulType different from MacWhisper?”

MacWhisper is a great transcription app, but it’s focused on file transcription, not real-time dictation. VaulType is designed for live dictation into any app with LLM post-processing and voice commands.


See CONTRIBUTING.md for the full guide. The quickest ways to contribute:

  • Report bugs via GitHub Issues
  • Add new processing modes (create a prompt template + register it)
  • Add new voice commands
  • Improve documentation
  • Test on different hardware and report compatibility
  • UI: Swift 5.9+ / SwiftUI / AppKit
  • Speech: whisper.cpp (C/C++) with Metal acceleration
  • LLM: llama.cpp (C/C++) with Metal acceleration
  • Audio: AVAudioEngine
  • Injection: CGEvent API
  • Storage: SwiftData
  • Build: Xcode 15+ / SPM / CMake
  • CI/CD: GitHub Actions

See TECH_STACK.md for detailed rationale.

PhaseFocusStatus
v0.1.0Menu bar + whisper.cpp + text injectionIn development
v0.2.0LLM post-processing modesPlanned
v0.3.0App-aware context, history, overlayPlanned
v0.4.0Voice commandsPlanned
v0.5.0Power user features, polishPlanned
v1.0Feature-complete stable releaseFuture

See ROADMAP.md for the full roadmap.

Yes:

Terminal window
git clone https://github.com/vaultype/vaultype.git
cd vaultype
./scripts/build-deps.sh
open VaulType.xcodeproj
# Build and run (⌘R)

See SETUP_GUIDE.md for detailed instructions.