VoiceStudio: local voice cloning, dubbing and an MCP server for your video agent

VoiceStudio isn't new (the repo dates to April), but it had a big weekend. @trending_repos made it repo of the day on Oct 17 with 2,135 stars in 24 hours, @RoundtableSpace posted it twice, and a wave of YouTube explainers in several languages followed. It had about 77,000 stars when we checked today. The pitch is a fully local alternative to ElevenLabs. You can clone a voice or design one from a description, dub videos with timed speech, dictate, transcribe, and make stories or audiobooks in batch, all on your own machine. Remote services are optional.

The ReelStack angle is the MCP server. It mounts on the running app at /mcp, so once VoiceStudio is open, an agent like Claude Code or Cursor can call generate_speech, clone_voice, design_voice, describe_voice, transcribe and list_voices. One practical tip from the docs: by default audio comes back to the agent as base64 inside the chat, which eats context. Set OMNIVOICE_MCP_OUTPUT_MODE=files and OMNIVOICE_MCP_BASE_PATH to a folder, and renders are written to disk so your HyperFrames or Remotion step can pick up the WAV (or Ogg/Opus, which needs ffmpeg) by path. The base path is also the security boundary for files the agent can read or write. The first generate call downloads the default OmniVoice model, about 2.3 GB, so install it ahead of time. There's also an agent skill (npx skills add debpalash/VoiceStudio).

Hardware, from the README: CUDA on NVIDIA (Windows/Linux), Metal on Apple Silicon, and CPU-only PCs work, just slower. Intel Macs get the UI only and have to connect to a remote backend. The latest release, v0.5.6 (Sep 23), fixed MCP inside Docker and added copyable setup for Claude Code, Cursor and Codex. v0.5.5 added cross-language cloning, so an English reference can read a French script on a multilingual engine.

The catch, which Stackdiff's video also flags: the app is AGPL-3.0, but model weights carry their own terms. VoiceStudio's own model-credits page says the default OmniVoice card pairs Apache-2.0 code with CC-BY-NC weights, and NLLB-200 is CC BY-NC as well. If you're dubbing client work, pick an engine whose weights allow commercial use and check the Model Manager's licence notes. The README is also clear that you should only clone voices you have permission to use. Spotted, not tested.