Voicebox is a free, open source desktop app (MIT license) that does on your machine what ElevenLabs and Wispr Flow do in the cloud: clones voices with AI from 10 to 30 seconds of audio, converts text to speech, transcribes your voice dictation, and thanks to its integrated MCP server, makes Claude Code, Cursor, or Windsurf speak to you with a voice that’s yours.
It’s also a project with 54,100 stars on GitHub and no release since April. Both things matter if you’re going to install it, so this guide covers what it does, how to connect it to your agent, and what its maintenance status means for you.
What is Voicebox?
Voicebox is a local-first AI voice studio created by Jamie Pine, the author of Spacedrive. The project presents itself as the piece that covers both halves of the voice cycle: output (text to speech and voice cloning, ElevenLabs territory) and input (dictation across the system, Wispr Flow territory), with a local LLM in the middle.
What the current release, v0.5.0, includes:
- Voice cloning and text to speech with seven engines: Qwen3-TTS, Qwen CustomVoice, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, HumeAI TADA, and Kokoro, interchangeable on each generation. Some clone from a sample; Kokoro and Qwen CustomVoice offer more than 50 predefined voices.
- Voice dictation with a global shortcut (hold down or toggle), transcribed with OpenAI Whisper and, if you want, cleaned of filler words by the local LLM.
- Captures: each dictation and recording is saved with its audio and transcription, and you can re-transcribe it or turn it into a voice sample.
- Stories Editor: a multitrack timeline for podcasts, dialogues, and narration.
- Post-production effects (pitch shift, reverb, compression, and more) built on Spotify’s
pedalboard. - Voice Personalities: you assign a free-form personality to a profile and a local Qwen3 model (0.6B, 1.7B, or 4B) rewrites the text in that character before reading it.
- A REST API and an MCP server at
127.0.0.1:17493.
The desktop app is built with Tauri (Rust), the backend with FastAPI (Python), and inference runs on MLX for Apple Silicon or PyTorch for the rest (CUDA, ROCm, DirectML, Intel XPU, or CPU). At the time of publishing this note, the repository has 54,100 stars and around 6,800 forks.
Is it a free alternative to ElevenLabs?
Yes, for cloning voices, generating audio, and having an API you can automate, with one condition: you provide the computing. As an alternative to ElevenLabs there’s no charge per character and the audio stays on your machine, but speed and quality depend on your hardware and the engine you choose. Anyone looking for “free ElevenLabs” finds a real answer here, not a trial plan.
The engines are not interchangeable, and choosing well is half the work:
| Engine | Languages | Strengths (according to the project) |
|---|---|---|
| Qwen3-TTS (0.6B / 1.7B) | 10 | Multilingual cloning, intonation instructions like “speak slowly” |
| Qwen CustomVoice | 10 | 9 predefined voices with natural language intonation control |
| LuxTTS | English | ~1 GB of VRAM, output at 48 kHz |
| Chatterbox Multilingual | 23 | The largest language coverage |
| Chatterbox Turbo | English | Fast 350M model with emotion tags like [laugh] and [sigh] |
| TADA (1B / 3B) | 10 (3B) | Long and coherent audio |
| Kokoro | 8 | 82M model, 50 predefined voices, fast on CPU |
The performance figures in the README (LuxTTS “150x real-time on CPU”, TADA with “over 700 s” of coherent audio, MLX “4-5x faster”) are numbers from the project itself, not independent benchmarks.
Does free text to speech work in Spanish?
Yes: four engines generate audio in Spanish, something we verified in the project’s source code and not just in its marketing. They are Qwen3-TTS (and Qwen CustomVoice), Chatterbox Multilingual, TADA 3B, and Kokoro. Kokoro comes with three predefined Spanish voices (Dora, Alex, and Santa), so you can generate a free AI voice in Spanish without recording anything.
Two caveats:
- Emotion tags don’t work in Spanish. Only Chatterbox Turbo interprets tags like
[laugh], and that engine is English only. The other engines read them aloud as text. - Transcription uses Whisper. The roadmap mentions new STT engines (Parakeet v3 and Qwen3-ASR) with better quality outside English, but those are planned work, not published.
How do you clone a voice with free AI locally?
To clone a voice with Voicebox you need between 10 and 30 seconds of clear speech, which you can record inside the app or upload as a file. The documentation recommends WAV (also accepts MP3, M4A, and FLAC), a quiet room, a steady tone, and complete sentences. With that you create a voice profile, choose a cloning engine, and generate a test phrase. If the result doesn’t convince you, add more samples to the same profile.
If you don’t want to clone anyone, Kokoro’s predefined voices work just as well, for example for agent notifications.
One limitation worth being clear about: the project’s RESPONSIBLE_USE.md file clarifies that Voicebox cannot verify who a voice sample belongs to, and expressly prohibits impersonating identities, committing fraud, or bypassing voice authentication systems. Clone your own voice or voices you have permission to use.
Does it work as voice dictation, just like Wispr Flow?
Partly. Voice dictation works on macOS and Windows with a global shortcut: you hold down the combination, speak, and release. But automatic pasting into the active text field, which is what makes Wispr Flow useful, today only works on macOS. On Windows and Linux automatic pasting is on the roadmap; in the meantime, the transcription stays saved in the Captures tab.
How do you install Voicebox?
Download the installer for your platform from the project’s site: a DMG for macOS (with separate versions for Apple Silicon and Intel) or an MSI for Windows. All binaries are also on the GitHub releases page. If you prefer to run it as a server without a desktop interface, with a web UI, clone the repository and use Docker:
git clone https://github.com/jamiepine/voicebox.git
cd voicebox
docker compose up
The interface is available at http://localhost:17493. The default docker-compose.yml only listens on localhost.
What to expect on first startup:
- Models are downloaded when you use them for the first time, not during installation. They range from about 350 MB (Kokoro) to about 8 GB (TADA 3B); according to the documentation, most people start with Qwen 1.7B, which is about 3.5 GB.
- Requirements: 8 GB of RAM minimum, 16 GB recommended, and 5 to 10 GB of free disk space. It works on CPU, but quite a bit slower; for real-time use a GPU is recommended. If you want to know what your hardware can run before downloading models of several gigabytes, check our guide on what hardware you need to run local models.
- Your data is saved in
~/Library/Application Support/sh.voicebox.app/on macOS and in%APPDATA%/sh.voicebox.app/on Windows. You can move models to another location with theVOICEBOX_MODELS_DIRenvironment variable.
Does Voicebox work on Linux?
Not directly. At the time of publishing this note, there are no precompiled binaries for Linux; the README points to instructions for building from source code at voicebox.sh/linux-install. On Linux, Docker is the practical way forward, and the unpublished changelog includes fixes for AMD GPUs with ROCm, both in Docker and in native installation. Automatic dictation pasting is also not available on Linux.
How do I make Claude Code speak to me in your voice?
You need three things: Voicebox open, a voice profile (cloned or predefined), and the MCP server registered with your agent.
Register the MCP server in Claude Code
Claude Code’s documented syntax takes the server URL as a positional argument:
claude mcp add --transport http voicebox http://127.0.0.1:17493/mcp \
--header "X-Voicebox-Client-Id: claude-code"
Note: the Voicebox README shows a variant with a --url flag. Claude Code’s MCP documentation doesn’t include that flag, so use the form above. Add --scope user if you want the server available in all your projects. Then verify the connection with claude mcp list, or with /mcp within a session.
The X-Voicebox-Client-Id header is not a credential. It’s a label that tells Voicebox which agent is calling, so it can apply its configuration.
Assign a voice to each agent
In Voicebox → Settings → MCP you can associate Claude Code with one voice profile and Cursor with another, so you know which agent is speaking to you without looking at the screen. The same panel shows when each client last connected, which confirms that the installation worked.
What about Cursor, Windsurf, or other MCP clients?
Voicebox documentation provides this block for clients with HTTP support like Cursor, Windsurf, and VS Code:
{
"mcpServers": {
"voicebox": {
"url": "http://127.0.0.1:17493/mcp",
"headers": { "X-Voicebox-Client-Id": "cursor" }
}
}
}
Don’t paste it as-is into Claude Code’s .mcp.json: Claude Code treats an entry with url but without "type": "http" as a configuration error. The repository’s own .mcp.json does include that field.
For clients that only support stdio, the app includes a voicebox-mcp binary. On macOS it’s at /Applications/Voicebox.app/Contents/MacOS/voicebox-mcp; the documentation lists paths for Windows and Linux. Voicebox has to be open for that binary to connect.
What can the agent do with Voicebox?
The server exposes four tools: voicebox.speak, voicebox.transcribe, voicebox.list_captures, and voicebox.list_profiles. The one you’ll use is speak, which takes the text, an optional profile (if you don’t pass one, it uses the voice assigned to the agent), an optional engine, the language, and personality: true to rewrite the text with the profile’s personality before reading it. The agent calls it when it decides to, so tell it when you want to hear it; for example: “let me know with voicebox when the tests finish”.
Things that don’t speak MCP, like shell scripts or CI jobs, can use the same path via REST with POST /speak.
Every time an agent speaks, an indicator with the voice name appears on screen for the entire duration of playback. The project’s argument is that silent text-to-speech running in the background is a trust problem: with the indicator, nothing speaks from your machine without you seeing it.
Is it safe to connect Voicebox to your agents?
On a single-user machine, the design is reasonable; outside of that, it’s not ready yet. In version v0.5.0:
- The server listens on
127.0.0.1and has no authentication. Any local process can call it, and bearer token authentication is on the roadmap. If you ever expose Voicebox beyond loopback, you’re exposing an unauthenticated API. - File path transcription is only accepted from loopback, so that a Voicebox listening on
0.0.0.0doesn’t become a way to read arbitrary files. Still, on shared machines the documentation recommends sending audio in base64. - Consent is your responsibility, as explained in the cloning section.
If your team already connects multiple MCP servers to your agents, the underlying problem is broader than this tool; we analyze it in MCP Gateway: the next major security category for agents.
Is Voicebox abandoned?
There’s no official answer, but the data calls for caution. As of September 16, 2026, the latest release is still v0.5.0, published on April 25, and the most recent commit on the main branch is from July 27. An open issue, #1101, directly asks whether the project is abandoned; at the time this note was published it had no response from the maintainer.
The project’s own status document, last updated on July 2, recorded 402 open issues, 88 open pull requests, and 34,800 stars. Today the repository shows 520 issues, 173 pull requests, and 54,100 stars. Interest grew more than half in ten weeks while the review queue nearly doubled.
That doesn’t make v0.5.0 unusable: it’s a complete and documented release, and everything this guide explains is based on it. But it changes how you adopt it:
- Take v0.5.0 for what it is. Don’t plan based on what’s on the roadmap (automatic clipboard paste on Linux, MCP authentication, new STT engines).
- Keep it on loopback. The lack of authentication won’t be fixed soon.
- For personal workflows, the risk is low: local audio, local models, and MIT license. For something a team depends on, review the repository activity again before committing.
Is it worth installing Voicebox?
Yes, if you want your code agent to notify you that the build finished, with your own voice or with a predefined voice in Spanish, without sending audio to anyone. Today it’s the most complete way to do it, and once you have a profile, configuration is a single command. Just keep in mind that you’re installing a finished release from a project that lost momentum, not a tool in full evolution.