DICTAre

Voice layer for AI coding agents.

Speak to your agent. No window focus required. 100% local.

dictare.io · OpenVIP Protocol

Why dictare

Most voice tools (Wispr Flow, Superwhisper, etc.) simulate keystrokes — they type into whatever window has focus. Switch to your browser and your code gets your voice.

Dictare uses a protocol. Your agent listens via SSE and receives transcriptions regardless of window focus. Your coding agent can be behind 3 other windows — it still gets your words.

Features

• No focus required — agent receives voice even when its window is in the background
• Agent-native — transcriptions go to the agent protocol, not a text field
• 100% local — STT runs on-device, zero data leaves your machine
• Multi-agent — switch agents by voice: "agent coding", "agent review"
• Open protocol — OpenVIP — any tool can implement the SSE endpoint
• Bidirectional — STT (voice in) + TTS (voice out)

Install

macOS:

git clone https://github.com/dragfly/dictare && cd dictare
./scripts/install.sh

Linux:

pip install dictare

Quick Start

# 1. Install as system service (auto-starts at login)
dictare service install

# 2. Connect your agent
dictare agent myproject --type coding

The service starts automatically. Speak — your agent receives the transcription.

How It Works

  Microphone
      │
      ▼
  STT Module       Whisper (MLX / CTranslate2) or Parakeet (ONNX)
      │             all local, zero cold-start
      ▼
  Pipeline         submit detection, agent switching, language filter
      │
      ▼
  OpenVIP          HTTP / SSE — open protocol
      │
      ▼
  Agent            receives transcription, no window focus needed

The engine runs as a background service (launchd on macOS, systemd on Linux). STT models are preloaded at startup. Each agent connects in its own terminal.

Agent Templates

Define agent types in ~/.config/dictare/config.toml:

[agent_types.coding]
command = ["claude"]
description = "AI coding assistant"

[agent_types.review]
command = ["aider", "--model", "claude-sonnet-4-6"]
description = "Code review"

[agent_types.writing]
command = ["claude", "--model", "claude-opus-4-6"]
description = "Writing and documentation"

Then connect using --type:

dictare agent myproject --type coding     # session "myproject", type "coding"
dictare agent frontend --type review      # session "frontend", type "review"
dictare agent -- claude --model opus      # explicit command override

Voice Commands

Submit triggers are multilingual (en, it, es, de, fr) and fully configurable.

Hotkey Cheat Sheet

Default hotkey: Right ⌘ (macOS) / Scroll Lock (Linux).

Service Management

dictare service install     # Install + enable (auto-starts at login)
dictare service start       # Start the service
dictare service stop        # Stop the service
dictare service restart     # Restart the service
dictare service status      # Show service and engine status
dictare service logs        # View recent logs
dictare service uninstall   # Remove the service

Keyboard Mode

No agent? Use dictare as a dictation tool — voice to keystrokes in any app.

dictare config set output.mode keyboard

Hotkey to toggle listening (configurable):

• macOS: Right ⌘ by default
• Linux: Scroll Lock by default

dictare config set hotkey.key KEY_RIGHTALT   # change hotkey

Text-to-Speech

dictare speak "Hello world"
dictare speak --engine piper "Hello"
echo "Hello" | dictare speak

Engines: espeak, say (macOS), piper, kokoro

Configuration

dictare config edit           # Open config in editor
dictare config list           # Show all settings
dictare config get stt.model
dictare config set stt.language it

Requirements

• Python 3.11
• macOS or Linux

macOS: Grant Input Monitoring permission when prompted during dictare service install. System Settings → Privacy & Security → Input Monitoring → enable Dictare.

Linux: Join input group: sudo usermod -aG input $USER (log out/in).

Development

git clone https://github.com/dragfly/dictare && cd dictare

# macOS Apple Silicon (MLX GPU acceleration)
uv sync --python 3.11 --extra mlx

# macOS Intel / Linux
uv sync --python 3.11

# Run engine in foreground
uv run --python 3.11 dictare serve

# Tests
uv run --python 3.11 pytest tests/ -x

# Tests (parallel)
uv run --python 3.11 pytest tests/ -x -n auto

Ghostty users: add keybind = shift+enter=text:\n to config. See TERMINAL_COMPATIBILITY.md.

Roadmap

• Plugin architecture: pipeline filters loadable as plugins, each declaring its model dependencies (STT, TTS, LLM, Vision).
• Realtime partial transcription: stream partial results while speaking using a fast small model.
• Cloud relay (Phase 2): E2E encrypted relay connecting web clients to local engines.

Protocol

dictare is the reference implementation of OpenVIP — an open protocol for voice input to AI agents. Any tool can implement the SSE endpoint and receive voice transcriptions from dictare.

License

MIT