July 20, 2026ToolMCPOpen Source

Voicebox: the Local Voice Studio Your Agent Can Speak Through

jamiepine/voicebox is climbing GitHub Trending with 629 stars added today, sitting at 43.3k total. It is an open-source, local-first AI voice studio — the pitch is ElevenLabs plus WisprFlow in one app, with models and data never leaving your machine. MIT licensed. Voice cloning from a sample or 50-plus presets, seven TTS engines including Qwen3-TTS and Kokoro, global dictation with push-to-talk hotkeys, Whisper speech-to-text, 23 languages.

The reason it belongs on this site is one feature: it ships an MCP server. Claude Code, Cursor or any MCP-capable agent can speak to you in a voice you cloned — including your own. The fleet-monitoring problem every agent operator has (five sessions running, which one needs me?) gets an ambient answer: the agent talks. Dictation the other direction runs through bundled local Qwen3 models that clean up your rambling into usable prompts, so the whole voice loop — you to agent, agent to you — stays on-device.

Honest caveat: the latest release, v0.5.0, is from April. The star surge is happening now, likely riding the local-first wave and the agent-voice use case. Trending is trending; the standing rule here is that a repo doing 600 stars a day earns coverage on signal alone.

Repo: https://github.com/jamiepine/voicebox
← Previous
From Pixels to States: World Models Should Think Like Game Engines
Next →
Super User Daily: July 20, 2026
← Back to all articles

Comments

Loading...
>_