The short answer
A voice layer lets the agent talk to you, not only you to it.
Claude Code’s /voice covers one direction: you speak and it transcribes your prompt. A voice layer adds the other. VoiceLayer does it from inside the agent: it is an MCP server with two tools, voice_speak and voice_ask, so Claude chooses when to talk and can ask you a question out loud. Heard does it from outside: it listens to Claude Code and Codex hooks and decides what to say, with no tools for the agent to call.
Inside means the agent controls the words and can hold a spoken back-and-forth. Outside means consistent updates in every session with no prompt changes, and one voice across several agents.
Side by side
Four ways to add voice to Claude Code.
From Claude Code’s docs and each project’s README. Checked October 6, 2026.
| Option | Direction | Who decides when it speaks | Engines | Platforms | Price |
|---|---|---|---|---|---|
| Claude Code /voicebuilt in | You to Claude only | No speech output | Transcribed on Anthropic’s servers | Where Claude Code runs locally, with a claude.ai account | Included |
| VoiceLayerMCP server | Both | The agent, by calling voice_speak or voice_ask | whisper.cpp local; edge-tts (online) | macOS, Linux | Free, Apache-2.0 |
| VoiceModeMCP + plugin | Both | The agent, in a conversation | Local Whisper.cpp and Kokoro, or OpenAI | macOS, Linux, Windows | Free, MIT |
| Heardhooks | Agent to you; you to agent on Power | Heard, from agent events | Kokoro, your own keys, or managed voices | macOS 14+ | Free engine; app from $15/mo |
VoiceLayer: speaking and listening tools for the agent.
VoiceLayer runs a persistent local daemon and exposes voice_speak for non-blocking speech and voice_ask for a spoken question that waits for your answer. It transcribes with whisper.cpp on your machine and speaks with edge-tts, Microsoft’s online voice service. On macOS it also ships VoiceBar, a small notch app. It installs from a Homebrew tap or npm, and it is new and actively updated.
brew tap etanhey/layers
brew install etanhey/layers/voicelayer
voicelayer setupHeard: a voice layer outside the agent.
Heard installs hooks for Claude Code and Codex and turns their events into short spoken updates. Because it sits outside the agent, the agent spends no tokens or tool calls on talking, every session behaves the same, and updates from several sessions share one queue with the project name first. Listening modes set how much you hear. On Power you reply by holding right ⌘, and Parrot answers questions about what your agents did.
Pick an MCP voice tool for spoken back-and-forth.
If you want Claude itself to choose its words and stop to ask you something out loud in the middle of a task, an MCP tool is the right shape. VoiceLayer is free, open source and runs on Linux as well as macOS. VoiceMode is the more established option of this kind and also runs on Windows.
Pick Heard to follow several agents without changing them.
If you run more than one session, or Codex next to Claude Code, a layer outside the agents keeps updates consistent and compressed, needs no instructions in your prompts, and keeps working when the agent forgets to call a tool.
Questions
Frequently asked questions
What is a voice layer for Claude Code?
Something that lets Claude Code talk to you, not only you to it. It can be an MCP server the agent calls, such as VoiceLayer or VoiceMode, or a layer outside the agent that speaks from its hooks, such as Heard.
Is Claude Code’s /voice a voice layer?
It is the input half. /voice transcribes your speech into the prompt; Claude Code has no built-in spoken output.
Is VoiceLayer fully local?
Its speech recognition is local whisper.cpp. Its speech output uses edge-tts, which calls Microsoft Edge’s online text-to-speech service.
Can I use VoiceLayer and Heard together?
Yes, though both may speak about the same moment. A common split is Heard for status updates and an MCP tool for spoken questions you answer.
Ready when your agents are
Hear the work without watching every screen.
Start with the full Heard Power experience for 14 days. No card required. The open source engine remains free.
Download Heard for Mac macOS 14 or later · Claude Code and Codex supported