debpalash / VoiceStudio

Install with prompt· agent preflight

Done is a clip that plays.

VoiceStudio ships an install guide written for coding agents. Paste one prompt, and your agent inspects the device, asks before every model download, and proves the install by generating a short clip.

Complete the setup, not just a plan. — the agent guide

Read the agent guide Download a release

Paste into your coding agentREADME

Install the VoiceStudio Electron app on this device and verify it works, following
https://github.com/debpalash/VoiceStudio/blob/main/docs/install/agent.md

Agents that support skills can also run npx skills add debpalash/VoiceStudio.

Six steps, in order.

The guide's procedure. The agent completes it and reports back, rather than handing you a to-do list.

  1. 01Load contextThe install guide for this OS, the performance notes and the voicestudio skill.
  2. 02InspectOS, CPU architecture, GPU, RAM and VRAM, free disk, and any existing install or models.
  3. 03InstallThe latest VoiceStudio-Electron release for this OS and architecture.
  4. 04ConfigureA voice-cloning engine and acceleration that fit this device.
  5. 05VerifyHealth, models, consent, install, then a real clip.
  6. 06ReportWhat was installed, where it lives and how to reopen it.

Step 05 · verify

Every rung needs an answer before the next.

The backend listens on http://localhost:3900 by default. The agent reads the API, asks before downloading, and polls until the model is really there.

One rung looks finished before it is. An empty jobs array counts as done only once GET /models reports "installed": true.

  1. 1GET /healthThe backend answers.gate
  2. 2GET /openapi.jsonDiscover the API instead of guessing it.read
  3. 3GET /models → "installed": falseState the model's size and license, and get consent.ask
  4. 4POST /models/installWith that entry's repo_id.act
  5. 5GET /models/install/statusPoll; stop on a failed state.poll
  6. 6"jobs": [] is not done yetFinished only once GET /models reports "installed": true.caution
  7. 7Generate a short clipWith a bundled or authorized voice, and confirm the audio file plays.done
Step 02 · compute

Check the actual device. Never assume a GPU.

NVIDIA GPU, Windows or LinuxCUDA acceleration
Apple SiliconMetal (MPS) acceleration
No dedicated GPUCPU, slower; a small CPU build of PyTorch, about 5 GB free disk
Windows on ARMExperimental: native ARM64 app, x64 Python backend under emulation, CPU only
Intel MacApp UI only; connect to a remote backend
Rules · from the guide

Limits the agent works within

  1. Install the Electron app. Tauri is archived; never install or launch it.
  2. Reuse existing data, settings and models.
  3. Ask before downloading a model, and state its size and license.
  4. Keep cloud services and analytics opt-in. Leave the analytics consent prompt to you.
  5. Name any permission or manual step it cannot perform instead of skipping it.

Step 06: the debrief you get back.

Six fields from the guide, left blank here because only your device can fill them in.

Version
Engine
Actual device
Data location
Audio output
How to reopen