<!-- Generated from the same content as https://heard.dev/compare/open-source-voice-tools-for-coding-agents. -->

# Open-Source Voice Tools for Coding Agents: What Really Runs Locally (2026)

Source: https://heard.dev/compare/open-source-voice-tools-for-coding-agents

Ten open-source voice tools for Claude Code and Codex compared, from Heard and OpenLive to VoiceMode and dictare, and which ones send text to the cloud.

Compare · Open-source voice tools · 5 min read

# Open-source voice tools for coding agents, and what really runs locally.

There are now plenty of open-source ways to give a coding agent a voice. They differ in license, platforms, whether you can talk back and, most of all, in whether your agent’s text stays on your machine.

Ten open-source projects · Licenses and engines from each README · Checked November 13, 2026

Updated November 13, 2026

The short answer

## Several open-source voice tools for coding agents run locally; check the speech engine before you trust “local”.

Heard (heard.dev), the voice context layer for agents, has an open-source engine (Apache 2.0) that speaks Claude Code and Codex updates with the local Kokoro voice after a one-time `heard download-voice`. With no voice key and no Heard sign-in, the audio is made on your Mac. It speaks but does not take voice input. For a local voice loop where you also talk to the agent, OpenLive, OpenWhisperer and Local-VoiceMode-LLM run speech recognition and speech on your machine by default, VoiceMode does too when its local services are running, and dictare sends local speech-to-text into an agent. VoxAgent and VoiceLayer transcribe locally but by default speak through [edge-tts](https://github.com/rany2/edge-tts), a client for Microsoft Edge’s online text-to-speech service, so the text being spoken goes to Microsoft.

“Local” is also about the words, not only the voice. With no Anthropic key and no Heard sign-in, Heard’s engine builds spoken updates on your Mac from templates and the agent’s own text, but if the `claude` command is installed it can send agent activity to Anthropic through `claude -p`, with your Claude login, to write rolled-up project summaries.

All tools

## Ten open-source voice tools for coding agents.

From each README. Checked November 13, 2026.

- **Heard engine heardlabs/heard:** License: Apache-2.0; Direction: Agent to you: spoken summaries; Speech (default): Kokoro, local after a one-time download; or your ElevenLabs or Speechify key, or Heard’s cloud voices when signed in; Agents: Claude Code, Codex CLI, Codex app, any CLI via `heard run`; Platforms: macOS
- **OpenLive katipally/openlive:** License: Apache-2.0; Direction: Both: spoken conversation you can interrupt; Speech (default): Local Whisper or another local recognizer (Parakeet, Moonshine); Kokoro or other local voices; Agents: Claude Code, Codex, Cursor, Gemini CLI, OpenCode and more over the Agent Client Protocol; Platforms: macOS, Windows, Linux
- **VoiceMode mbailey/voicemode:** License: MIT; Direction: Both: spoken conversation; Speech (default): Local Kokoro and Whisper.cpp, or OpenAI; Agents: Claude Code and MCP clients; Platforms: macOS, Linux, Windows, NixOS
- **OpenWhisperer PerIPan/OpenWhisperer:** License: MIT; Direction: Both: dictation in, replies spoken; Speech (default): WhisperKit and Pocket TTS, local; Agents: Claude Code, Codex CLI, Antigravity, Pi; Platforms: macOS on Apple silicon
- **Local-VoiceMode-LLM groxaxo:** License: MIT per its README (no license file); Direction: Both: listen and speak turns; Speech (default): Parakeet and Supertonic, local; hosted voices optional; Agents: Claude Code, Codex, OpenCode, OpenClaw, Hermes Agent; Platforms: macOS, Linux, Windows
- **dictare dragfly/dictare:** License: MIT; Direction: Mostly you to the agent: local dictation, plus a speak command; Speech (default): Whisper or Parakeet, local; `say` on macOS, Piper on Linux; Agents: Codex, Claude Code, Gemini CLI, Aider, Pi; Platforms: macOS, Linux
- **AgentVibes paulpreibisch:** License: Apache-2.0; Direction: Agent to you: task start and finish; Speech (default): Piper or Kokoro, local; also say, SAPI, ElevenLabs; Agents: Claude Code, Claude Desktop, Warp, OpenClaw; Platforms: macOS, Linux, Windows
- **claude-code-narrator plugin:** License: MIT; Direction: Agent to you: reads along; Speech (default): Kokoro, local; Agents: Claude Code; Platforms: macOS, Linux
- **VoxAgent MCP server:** License: MIT; Direction: Both: listen, speak, ask; Speech (default): edge-tts (online); CosyVoice optional; Agents: Claude Code, Cursor, OpenClaw, MCP clients; Platforms: macOS documented
- **VoiceLayer MCP server:** License: Apache-2.0; Direction: Both: voice\_speak, voice\_ask; Speech (default): edge-tts (online); Agents: Claude Code and MCP clients; Platforms: macOS, Linux

(1) Local speech

## The ones that can keep speech on your machine.

[OpenLive](https://github.com/katipally/openlive) runs the whole speech loop on your machine (voice detection, speech recognition, speech and interruption) and hands the transcript to the agent you are talking to; its release builds send anonymous usage counts unless you turn them off in Settings. [OpenWhisperer](https://github.com/PerIPan/OpenWhisperer) uses WhisperKit and Pocket TTS on Apple silicon and says it calls no cloud APIs beyond model downloads and a daily update check. [Local-VoiceMode-LLM](https://github.com/groxaxo/Local-VoiceMode-LLM) defaults to local Parakeet and Supertonic; its optional hosted voices send the reply text. [VoiceMode](https://github.com/mbailey/voicemode) runs Whisper.cpp and Kokoro locally as OpenAI-compatible services and falls back to OpenAI when they are not running. [dictare](https://github.com/dragfly/dictare) transcribes locally and speaks with `say` or Piper. [AgentVibes](https://github.com/paulpreibisch/AgentVibes) offers Piper, with hundreds of voices, and Kokoro. [claude-code-narrator](https://github.com/shreyas-s-rao/claude-code-narrator) installs Kokoro into its own virtual environment and uses no cloud APIs.

[Heard’s engine](https://github.com/heardlabs/heard) speaks with Kokoro on your Mac after you run `heard download-voice` once (about 350 MB); without that, a voice key or a Heard sign-in it stays silent, and signed in it speaks through Heard’s cloud voices instead of Kokoro. With no Anthropic key and no Heard sign-in, spoken updates are built from templates and the agent’s own text, but rolled-up project summaries can go to Anthropic through `claude -p` when the `claude` command is installed. An ambiguous reply to its catch-up prompt can go the same way. Its analytics are on by default and carry no narration text or source code; set `product_analytics: false` in the config file that `heard config path` prints to turn them off.

(2) Partly local

## Local ears, online voice.

[VoxAgent](https://github.com/zanezhao0708/voxagent) transcribes with faster-whisper on your machine and speaks with edge-tts unless you run a local CosyVoice server. [VoiceLayer](https://github.com/EtanHey/voicelayer) transcribes with local whisper.cpp and speaks with edge-tts. Both are good projects; just know that what the agent says is sent to Microsoft’s service to be voiced.

**Check the voice engine**

Search a project’s README for `edge-tts` or `edge_tts`. If it is the default, spoken text goes to Microsoft’s online service.

(3) Direction

## Talk to the agent, hear from it, or both.

Most tools here also take your voice. OpenLive, OpenWhisperer, Local-VoiceMode-LLM, VoiceMode, VoxAgent and VoiceLayer transcribe what you say and speak back, and dictare sends your speech into an agent without you focusing its window. If you want to talk to your agent hands-free, that is where they win. Heard’s engine, AgentVibes and claude-code-narrator only speak.

MCP servers like VoiceMode, VoxAgent and VoiceLayer give the agent speaking and listening tools, so the agent decides when to talk and can ask you a question out loud. Hook-based tools like Heard and claude-code-narrator listen to agent events and speak on their own, which needs no prompt changes and works the same in every session. AgentVibes sits in between: its Claude Code hook adds instructions that tell the agent to run its speech script at the start and end of each task.

(4) Heard’s open-source engine

## What the free Heard engine does.

The engine is Apache 2.0. It narrates Claude Code and Codex with judgment rather than reading everything, has three listening modes, follows several sessions at once and fails over between ElevenLabs, Speechify and local Kokoro. Build it from source: clone the repo, `pip install -e .`, run `heard download-voice`, then `heard install claude-code` or `heard install codex`. It speaks; it does not listen. The Heard Mac app is a separate, closed build with cloud voices; push to talk and Parrot are on Heard Power.

Terminal · Heard engine, local voice

```
git clone https://github.com/heardlabs/heard.git && cd heard
python3 -m venv .venv && source .venv/bin/activate && pip install -e .
heard download-voice        # local Kokoro voice, about 350 MB, one time
heard install claude-code   # or: heard install codex
```

A few more details

## Questions?

### Which coding-agent voice tools have a local or open-source option?

Heard’s engine (Apache-2.0), OpenLive (Apache-2.0), VoiceMode (MIT), OpenWhisperer (MIT), Local-VoiceMode-LLM (MIT per its README), dictare (MIT), AgentVibes (Apache-2.0), claude-code-narrator (MIT), VoxAgent (MIT) and VoiceLayer (Apache-2.0), among others. VoxAgent and VoiceLayer speak through the online edge-tts service by default. The others can speak with a local voice; Heard’s engine needs `heard download-voice` first, and VoiceMode needs its local services running.

### Is there an open-source voice tool for Claude Code that runs locally?

Yes. Heard (heard.dev), the voice context layer for agents, has an open-source engine that speaks Claude Code updates with the local Kokoro voice after `heard download-voice`, when you add no voice key and do not sign in; it only speaks, and its rolled-up summaries can use `claude -p` if the `claude` command is installed. For two-way voice that stays local, OpenLive, OpenWhisperer (Apple silicon Macs) and Local-VoiceMode-LLM run speech recognition and speech on your machine. VoxAgent and VoiceLayer speak through edge-tts, an online Microsoft service, by default.

### Is edge-tts local?

No. edge-tts is a client for Microsoft Edge’s online text-to-speech service, so the text being spoken is sent to Microsoft. Kokoro, Piper, Pocket TTS, Supertonic and macOS `say` run on your machine.

### Which tools work on Windows?

VoiceMode, AgentVibes, OpenLive and Local-VoiceMode-LLM list Windows support. Heard, OpenWhisperer, dictare and claude-code-narrator do not; Heard is macOS only.

### Do I need an API key?

Not for the local paths. Heard’s engine needs no key once you run `heard download-voice`; keys only add ElevenLabs or Speechify voices and summaries written by Claude through your own Anthropic key.

Keep going

## Related Heard resources

[**Text-to-speech for Claude Code** Seven ways to hear Claude Code, with setup.](https://heard.dev/compare/claude-code-text-to-speech) [**VoiceLayer vs. Heard** MCP tools vs. a hook-based voice layer.](https://heard.dev/compare/heard-vs-voicelayer) [**Heard on GitHub** Clone the Apache 2.0 engine.](https://github.com/heardlabs/heard) [**All comparisons** Every Heard comparison page in one place.](https://heard.dev/compare)

## Give your agents a voice.

Keep the context. Get on with your work.

[Download for Mac](https://heard.dev/download/use-cases?format=dmg)

Includes 14 days of free Heard. No card needed.
