VoiceStudio: The Free, Local ElevenLabs Alternative, Checked

On October 16, 2026 a Spanish-language post by @Nozelcode went around X: a solo developer has open-sourced a "100% free" replacement for ElevenLabs that runs entirely on your own machine, the GitHub repo is at 19.4K stars, and it can clone a voice from a clean reference clip.

The project is VoiceStudio. We checked the post against GitHub's API, the repository's own documents and the model it runs on. The app is real and genuinely free to use. The post's star count is six weeks old, and "100% free" skips one important detail: the voice model it starts with is licensed for non-commercial use only.

The original post by @Nozelcode on X (October 16, 2026)

What VoiceStudio is

A desktop app (Electron, for macOS, Windows and Linux, plus Docker) with a Python backend that runs speech models on your own hardware. Its README lists six jobs:

The default speech engine is OmniVoice from k2-fsa, and the app can switch to more than a dozen others, including CosyVoice 3, IndexTTS 2.5, KittenTTS and MLX-Audio. Hardware support, from the README:

What the post gets right, and what it does not

The star chart

The repository was created on April 9, 2026. For five months it grew steadily; then September brought almost 38,000 new stars, with the busiest single day on September 28 (4,331 stars).

What it replaces, in our catalog

VoiceStudio is now listed in AgentGid's voiceover catalog, next to the paid services and the other open models. The question that matters for a creator is not "is it free" but "may I publish what it makes":

For a sense of scale: a 500,000-character manuscript costs $40 at ElevenLabs' v3 API rate, by our calculation (500,000 ÷ 1,000 × $0.08). VoiceStudio does it for the price of your electricity, but with the default model you may not sell the result. Inside the app you can switch to an engine whose license allows it; its Model Manager shows each model's source notices, though the project's license notice says it asserts no commercial clearance for any model.

Our Short-video kit makes the same point in its voiceover step: "several free voices forbid commercial use", so it starts with Kokoro, Chatterbox and Piper.

Use it from your coding agent

This is the part the post does not mention, and the most interesting one for AgentGid readers. VoiceStudio's backend has an MCP server built in, so an agent can speak, clone and transcribe locally while the app is open:

The endpoint is http://localhost:3900/mcp/ (keep the trailing slash). The app exports ready configs for Claude Code, Cursor and Codex under Integrations. For Codex, the docs give this entry for ~/.codex/config.toml:

[mcp_servers.voicestudio]
url = "http://127.0.0.1:3900/mcp/"
http_headers = { "X-OmniVoice-Client-Id" = "codex-cli" }

The header names the agent, so each agent can be bound to its own voice. Two details from the docs save a failed first run:

• Install the model first. On a fresh machine the first generate_speech call also downloads OmniVoice, about 2.3 GB, and can time out. Install it from the app's first-run setup or Model Catalogue.
• Keep audio out of the chat. By default the tool returns the WAV as base64 inside the conversation, which costs tokens. Set OMNIVOICE_MCP_OUTPUT_MODE=files and OMNIVOICE_MCP_BASE_PATH to a shared folder, and the agent gets a file path instead.

Scripts that only need transcription can skip MCP: the backend answers OpenAI-style at POST http://127.0.0.1:3900/v1/audio/transcriptions.

To install, the README offers a one-line installer for macOS and Linux (curl -fsSL https://voicestudio.sh/install | sh), installers on the releases page (v0.5.6 is the latest, from September 23), and a prompt you can paste into your coding agent that points it at the project's agent install guide. As with any script piped into a shell, read it first.

The verdict

• Worth installing if you want voice cloning, dubbing or dictation on your own machine, or a voice for your coding agent through MCP. The app is free, and your audio stays local.
• Not a drop-in ElevenLabs replacement for paid work with its default model: OmniVoice's weights are non-commercial. Pick an engine with a commercial license, or keep a paid service for client work.
• Clone only voices you have permission to use. Both the project and the OmniVoice model card say so.

Every number above comes from GitHub's API, VoiceStudio's repository and site, the OmniVoice model card or our voiceover catalog, read on October 17, 2026. We did not run VoiceStudio ourselves.