AnythingLLM

magic-echo

2 min read

Requirements & Compatibility

  • OS Support: Windows (no special permissions), macOS (requires Accessibility and Screen Recording permissions). Linux is not supported.
  • AnythingLLM Version: Desktop v1.15.0 or later.
  • Hardware/Model Requirements: Smart Transcription with On-Screen Awareness requires a configured LLM provider supporting vision/multi-modal models.

Operating Modes

1. Quick Dictation (Press-and-Speak)

  • Start: Press activation shortcut (Default: Option+Z on macOS, Alt+Z on Windows).
  • Process: Speaks naturally; stops and auto-submits when silence is detected (configured via Silence Detection setting).
  • Cancel: Press Esc while active.
  • Force Submit: Press Enter to bypass silence detection.

2. Extended Dictation (Hold-to-Speak)

  • Start: Hold the activation shortcut.
  • Process: Continues recording until manually stopped.
  • Cancel: Press Esc to discard.
  • Submit: Click the “Stop” button in the widget.

Processing Modes

  • Smart Transcription: Transcribes audio, then processes via configured LLM for grammar, punctuation, formatting, and screen context. (Subject to daily limits on free tier; unlimited on local models/Pro). Falls back to Raw Transcription automatically if LLM processing fails.
  • Raw Transcription: Local on-device transcription without LLM processing. Faster, unlimited usage, skips grammar/punctuation cleanup.

Configuration Settings

Navigate to Settings -> Magic Echo to configure:

SettingDescription / Options
Activation KeyKey combined with Option (macOS) / Alt (Windows) to trigger dictation. Default: Z
Default Processing ModeSmart or Raw
On-Screen AwarenessEnable/Disable visual screen context for LLM. (Requires vision-compatible LLM)
Preferred MicrophoneSelect input device
Silence DetectionAggressive / Average / Relaxed (Controls auto-submit timing)
Widget SizeDefault / Large / Huge / Max
Voice CommandsConfigure static keyword triggers to paste predefined text blocks.
Custom VocabularyAdd comma-separated jargon, technical terms, or names (Smart mode only).

Voice Commands

Configure at Settings -> Magic Echo -> Voice Commands.

  • Directly maps a spoken exact phrase to a text snippet (bypasses LLM, unlimited usage).
  • Example: "PRD Template" -> pastes Markdown template; "sign off" -> pastes "Best regards,\n[Name]".

Troubleshooting & Performance Optimization

macOS Permissions Fix

If text insertion or screen capture fails:

  1. Quit AnythingLLM.
  2. Open macOS System Settings -> Privacy & Security.
  3. Manually add and enable AnythingLLM under both Accessibility and Screen Recording.
  4. Restart AnythingLLM.
  5. Disable and re-enable Magic Echo in settings.

Latency Reduction

  • Disable On-Screen Awareness to skip screenshot processing.
  • Set Default Processing Mode to Raw Transcription.
  • Note: The first dictation session after inactivity takes longer due to model loading; subsequent runs remain warm.

Graph View