Local dictation for
agentic coding workflows
Press a key and speak naturally. Clean, formatted text appears in Claude Code, Cursor, VS Code, or any terminal. Transcription and editing run on your device, never the cloud.
Drive Claude Code, Cursor, and terminals by voice
Prompt your AI coding assistants, run shell commands, write commit messages, and draft code comments without touching the keyboard. Evoglyph injects clean, formatted text into the focused field, choosing the right strategy for each app.
Native Cocoa editors
Xcode, TextEdit, Notes, and other native macOS apps receive text via AXUIElement read-splice-write. Text lands at the caret with no clipboard side effects and full undo history.
See Text injection for the full strategy table.
VS Code, Cursor & Electron apps
Electron-based editors and web-rendered surfaces receive text via Pasteboard + Cmd+V. Dismiss Cursor and VS Code autocomplete popups with Escape before dictating.
Per-app notes at Target apps.
Terminals (iTerm2 & others)
Terminal-class apps receive text via Pasteboard + Cmd+V. iTerm2's Secure Keyboard Entry blocks all injection; disable it in iTerm2 › Secure Keyboard Entry while using Evoglyph.
Details at Target apps.
The injection strategy is automatic; there's nothing to configure per app. For the full routing table and per-app notes, see Text injection and Target apps.
Developer vocabulary: gRPC,
OAuth,
Postgres
The base model is general-purpose. Add your tech stack's jargon once and Evoglyph biases the decode toward those terms across every dictation.
- CTC keyword boosting increases the probability of your terms winning the decode; no model retraining, no per-user fine-tuning
-
Handles acronyms, camelCase, and compound names:
WebSocket,TypeScript,SwiftUI,Kubernetes - Changes take effect on the next dictation; no restart, no apply step
- The vocabulary list and all scoring stay on-device; nothing leaves your Mac
Example: tech stack terms
-
gRPC
“grumpy C” →
gRPC -
Postgres
“prosegrass” →
Postgres -
OAuth
“oh auth” →
OAuth -
WebSocket
ambiguous →
WebSocket
Examples from the vocabulary docs.
Per-app setup
The Target apps page covers every supported app: the exact injection path, known gotchas, and notes for Claude desktop, iTerm2, VS Code, Cursor, Slack, and more.
View the target-apps guide →Fast enough to keep pace with thought
Evoglyph's speech-to-text engine-time, from end-of-speech to decoded text, measures 85 ms at the p50 on Apple Silicon. That's the transcription model alone; total time from hotkey release to text appearing includes cleanup latency.
For the full numbers, methodology notes, and accuracy results, see /benchmark.
85 ms
p50 engine-time
124 ms
p95 engine-time
2.41%
WER (LibriSpeech test-clean, our harness)
Engine-time: end-of-speech to decoded text, transcription model only. Figures from internal eval results, 2026-05-11; methodology available on request via [email protected].
Your code never leaves the machine
Every transcription and AI cleanup step runs on your Mac. Audio is captured, processed, and discarded locally; no audio, no text, no prompts are ever sent to any server.
- Speech is transcribed on-device by Parakeet TDT v2 on the Apple Neural Engine
- AI cleanup runs on-device; the cleanup model is downloaded once at first launch and stays local
- No analytics, no telemetry by default; crash reports are opt-in only
- Dictation history stays in a local SQLite database, never synced to a server
The network calls Evoglyph makes
- 1. Model download Hugging Face, first launch only.
- 2. License validation evoglyph.com licensing service, periodic re-check. No audio involved.
- 3. Free-trial start & usage sync Trial only: evoglyph.com licensing service. A pseudonymous device hash and the trial word count — never dictation content.
- 4. Update check Sparkle against evoglyph.com. Version numbers only.
- 5. Crash and error reports Evoglyph’s own servers (diag.evoglyph.com), opt-in only, off by default in Settings › Diagnostics. No third-party telemetry vendor.
Audio never leaves your device. See where-your-audio-goes.html for the full breakdown.
Start dictating into your dev tools
Requires Apple Silicon and macOS 14.2 (Sonoma) or later.