Skip to content

agent.md

Latest commit

 

History

History
44 lines (39 loc) · 2.66 KB

agent.md

File metadata and controls

44 lines (39 loc) · 2.66 KB
Raw

Install with a coding agent

Instructions for a coding agent (Claude Code, Codex, Cursor, …) asked to install VoiceStudio. The short prompt in the README points here. Complete the setup, not just a plan.

Rules

  • Install the Electron desktop app. Tauri is archived; never install or launch it.
  • Reuse existing data, settings and models. Ask before downloading a model: state its size and license.
  • Keep cloud services and analytics opt-in. Leave the first-run analytics consent prompt to the user; never answer it for them.
  • Verify the actual compute device; do not assume GPU support.
  • Name any permission or manual step you cannot perform instead of skipping it.

Steps

  1. Load context. Read the install guide for this OS (macOS · Windows · Linux; Docker is a headless backend, not the desktop app), performance and skills/voicestudio/SKILL.md. If your agent supports skills, install it with npx skills add debpalash/VoiceStudio.
  2. Inspect the device. OS, CPU architecture, GPU, RAM/VRAM, free disk, and any existing VoiceStudio install, backend or downloaded models.
  3. Install.
    • Default: the latest stable release asset named VoiceStudio-Electron-* for this OS and architecture (macOS .dmg, Windows .exe, Linux .AppImage or .deb), or the install script.
    • From source: follow electron/README.md — bun install, bun run setup:api, bun run dev from the repository root. Let Electron supervise the backend; do not start a second one.
    • Migrating from Tauri: back up first and follow the migration guide.
  4. Configure. Pick a supported voice-cloning engine and acceleration that fit this device. Keep working defaults and install required dependencies.
  5. Verify. Start the app and check /health at the backend address (default http://localhost:3900); discover the API through /openapi.json. If the chosen engine's model shows "installed": false in GET /models, state its size and license and get consent, then send POST /models/install with that entry's {"repo_id": "…"}. Poll GET /models/install/status: stop on a failed state, and treat an empty jobs array as finished only once GET /models reports "installed": true. Then generate a short clip with a bundled or authorized voice and confirm the audio file plays.
  6. Report. Installed version, engine, actual device, data location, audio output path, and how to reopen the app.