Larmor Open-Sourced: Local Voice Interface for MCP-Compatible AI Agents

Larmor, an open-source tool released under the MIT license, enables natural two-way voice interaction with AI coding agents. It addresses common issues in voice-controlled agents, such as echo cancellation, distinguishing interruptions from background noise, and handling input while the agent is processing.

Technical Implementation

The tool runs entirely on local models on Apple Silicon Macs, utilizing Parakeet for speech-to-text, Chatterbox for voice synthesis, and Apple's native echo cancellation. It integrates with any agent supporting the Model Context Protocol (MCP), including Claude Code, Codex, and Gemini CLI.

Key features include:

• Interruption Handling: The agent pauses voice output immediately when the user starts speaking, only stopping for clear speech rather than acknowledgments like "mm-hmm."
• Mid-Work Input: A hook mechanism passes user speech to the agent at its next tool call, allowing direction changes without returning to the keyboard.
• Voice Turn Management: A Stop hook ensures the agent continues listening after a voice turn, preventing premature session ends.

Installation and Usage

Larmor is installed via a curl command and requires an initial download of approximately 3.7 GB of models. In Claude Code, users activate it with the /larmor command. The project includes an MCP server with speak and listen tools, an installer, and a menu-bar application for managing the speech engine.