VoiceBox: Local Voice Cloning and MCP for AI Agents

VoiceBox, an open-source desktop voice application from developer Jamie Pine, reached v0.5.0 with 53,000 GitHub stars and 1.3 million downloads, bundling local voice cloning, a seven-engine TTS pipeline, Whisper dictation, and a built-in Model Context Protocol server at http://127.0.0.1:17493/mcp. The MCP server exposes four tools — voicebox.speak, voicebox.transcribe, voicebox.list_captures, and voicebox.list_profiles — that Claude Code, Cursor, Windsurf, and Cline can call, a local endpoint no cloud voice service offers. VoiceBox's cloning engine, Alibaba's Apache 2.0-licensed Qwen3-TTS, reports a 1.835% word error rate across ten languages and a 0.787 speaker similarity score, which the project says outperforms ElevenLabs Multilingual v2 and MiniMax on accuracy, while ElevenLabs charges $22 a month for voice cloning.

ElevenLabs charges $22 a month to clone your voice. VoiceBox does it for free, runs entirely on your machine, and lets your AI coding agents speak back to you in that voice. That last part is the one no cloud service can match. VoiceBox github.com/jamiepine/voicebox https://github.com/jamiepine/voicebox is an open-source desktop application that bundles voice cloning, a seven-engine TTS pipeline, Whisper dictation, and a built-in MCP server into one local install. It hit v0.5.0 this month with 53,000 GitHub stars and 1.3 million downloads. The pitch is direct: every voice feature you pay for in the cloud, running on your hardware, no data leaving your machine. Your Coding Agent Now Has a Voice The standout feature in v0.5.0 is the MCP server. VoiceBox runs a local Model Context Protocol https://modelcontextprotocol.io/ endpoint at http://127.0.0.1:17493/mcp , and it exposes four tools any MCP-aware client can call: voicebox.speak , voicebox.transcribe , voicebox.list captures , and voicebox.list profiles . Claude Code, Cursor, Windsurf, and Cline all support it. Connecting takes about four lines in your MCP config file and a restart: { "mcpServers": { "voicebox": { "command": "/Applications/Voicebox.app/Contents/MacOS/voicebox-mcp" } } } Once connected, an agent can call voicebox.speak with a text string and a voice profile name. VoiceBox plays the audio locally and surfaces an on-screen pill showing the profile name and a live elapsed timer — so you always know when your agent is speaking and in whose voice. Per-client binding lets you pin Claude Code to one cloned voice and Cursor to a different preset, entirely through the Settings → MCP panel. No cloud voice service offers a local MCP endpoint. This is the gap VoiceBox fills, and it is a real one as agentic workflows become standard in development. Voice Cloning That Can Beat ElevenLabs on Accuracy The cloning engine is Qwen3-TTS, the model Alibaba open-sourced in January 2026 under Apache 2.0 after training on five million hours of speech data. Its benchmarks are worth knowing: 1.835% word error rate across ten languages and a 0.787 speaker similarity score — numbers that outperform ElevenLabs Multilingual v2 https://elevenlabs.io/pricing and MiniMax on accuracy. In practice: record a short audio sample, import it into VoiceBox, and a cloned profile is ready in minutes. If you do not need cloning, there are 50+ preset voices and six other TTS engines to choose from — Kokoro, Chatterbox, Chatterbox Turbo, Luxtts, TADA, and the standard Qwen engine. Twenty-three languages are supported across the pipeline. Voice profiles can also carry a personality: a local LLM sits between the agent’s text output and the TTS engine, reshaping what gets spoken according to the profile’s configured style. An agent that writes terse commit messages can still speak in complete, conversational sentences if that is what the persona specifies. Dictation, Without the Subscription VoiceBox also replaces WisprFlow https://wisprflow.com/ . A global hotkey right Cmd + right Option on macOS by default, right Ctrl + right Shift on Windows captures your voice from any app, runs it through Whisper locally, optionally refines the transcript with a local LLM, and pastes the result into whatever text field was focused. An on-screen pill walks through each stage — recording, transcribing, refining, done — with a live timer. Everything happens on-device. No audio is uploaded. For developers who dictate into their IDE, Slack, or browser, this is a straightforward WisprFlow replacement that costs nothing and retains no data. Getting Started Download the installer from voicebox.sh https://voicebox.sh Mac, Windows, Linux . The app bundles its dependencies, so no separate Python or Bun install is required for end users. First run: create a voice profile clone or choose a preset , then go to Settings → MCP and copy the client-specific configuration snippet. The settings page pre-fills the correct absolute path to the voicebox-mcp binary for your platform. Full documentation, including the MCP server reference and dictation setup, is at docs.voicebox.sh https://docs.voicebox.sh . One caveat worth stating directly: VoiceBox is an MIT-licensed project maintained primarily by one developer. There is no commercial support, no SLA, and 400+ open GitHub issues that signal active but resource-constrained development. It is excellent tooling for individual developers and small teams. Enterprise deployments should weigh the bus-factor risk and, as always, obtain proper consent before cloning anyone’s voice. The Local-First Moment ElevenLabs’ recent announcement of an on-premise enterprise tier confirms the direction: local voice processing is becoming a serious requirement, not a niche preference. VoiceBox arrived at this moment as the open-source answer — combining the two tools most developers would pay for voice cloning and dictation with the agent integration layer those tools are missing. The MCP server in particular is the bet worth watching. As more coding agents ship MCP support by default, a local voice layer that any agent can call becomes infrastructure rather than a nice-to-have. VoiceBox is early, rough in places, and built by a small team. It is also the only tool doing all of this locally.