Pause
Freezes playback in place. Resume continues. Pause alone never opens the regular microphone.
Local voice for authorized MCP clients
Talk to your AI.
Interrupt naturally.
Keep the thread.
AI Voice MCP turns a typed agent loop into a real spoken exchange. Hear the answer, pause it, say a chosen interruption phrase, or take the floor with one click. The next response remembers what was being said and adapts to your new input.
- Windows x64 and ARM64
- Local speech recognition
- Apache-2.0
Works with MCP-capable clients
Codex
Claude
Gemini CLI
LM Studio
Local models
The two-listener loop
Your turn starts with a beep.
Your interruption starts in silence.
One shared microphone and one shared Whisper model move between two distinct listening modes. There is no competing second process and no lost response context.
-
01
The AI speaks
Playback starts. A no-beep phrase listener quietly watches only for words you selected.
-
02
You say “umm”
The default phrase pauses playback exactly where it is and yields the microphone.
-
03
The cue sounds
The regular full-transcription listener opens and plays the familiar three-beep “your turn” cue.
-
04
The conversation adapts
Silence resumes the same audio. New speech replaces the unheard remainder and returns the prior context so the AI can revise, not restart.
You own the floor
Four controls. Four clear meanings.
Interrupt / Skip
Discards unheard audio, sounds the listening cue, and gives you the floor immediately.
Your phrase
Say “umm” or choose another phrase. It triggers silently during playback or while held on Pause.
Stop
Ends playback and the exchange. Nothing keeps listening after you stop the session.
The context-preserving part
It does not forget the answer it was in the middle of giving.
When you add something new, Voice Command returns the unheard response, playback position, queued context, and your interruption to the connected AI. The AI can address your new point while preserving whatever still matters from the original answer.
PRIOR RESPONSE
“There are three routes. The first is fastest, while the second gives you…”
YOU INTERRUPT
“Umm—only compare the options that work offline.”
REVISED RESPONSE
“That narrows it to two. Here is the faster offline option first…”
Local boundary
Your microphone is not a website feature.
Microphone capture, faster-whisper transcription, silence detection, phrase matching, and playback control run on your Windows computer. The listener binds to localhost:5123, not your LAN or this public site.
Text-to-speech uses Edge TTS by default and therefore reaches Microsoft’s online TTS service. You can replace it with a local backend. Voice adds no new tool privileges; sensitive actions still use the permissions and confirmation boundaries of your AI client.
Install with your AI
Paste one request. Let your coding agent do the setup.
The repository supports Windows x64 and ARM64. Your agent should back up existing MCP configuration, install the local Voice App, wire the Voice MCP server, and tell you which manual reopen step remains.
PASTE TO YOUR CODING AGENT
Install AIWander/Voice-Command for this computer. Set up the local Voice App and Voice MCP, preserve and back up my current configuration, choose the correct Windows architecture, do not restart or close anything for me, and verify the install before telling me what I need to reopen manually.