deepakkrsoni

Voice Instinct — Development skill for Claude Code

Development community

Voice-first local assistant: talk to it, it delegates real work to your own Claude Code (or other agent) and asks you by voice when it needs a decision.

How to install Voice Instinct

This entry records only its repository, not the path inside it, so there is no exact command to give. Open deepakkrsoni/voice-instinct and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Voice Instinct does

Voice-first local assistant: talk to it, it delegates real work to your own Claude Code (or other agent) and asks you by voice when it needs a decision. Not affiliated with Instinct.

Alternatives in Development

Claude Code Voice Skill — Skill to talk to Claude about your projects over the phone 161 ★
Generate Assets — Generate real game assets — images, textures, skyboxes, 3D models, SFX, music, voice, and intro/cinematic vide 76 ★
Backtalk — Talk to your Claude Code agent out loud 37 ★

README

voice-instinct

A voice-first assistant that runs on your own machine. You talk to it; a fast "front" model handles the conversation and hands real tasks (files, shell, web, code) to a backend agent, then speaks the result. Not affiliated with Instinct.

mic -> Parakeet STT -> Smart Turn v3 -> front LLM -> Kokoro TTS -> speaker
                                          |  ^
                                 delegate(task)  result
                                          v  |
                              Claude Code CLI (your login)

Run

uv sync
echo 'GEMINI_API_KEY=...' > .env      # or whatever api_key_env names in config.toml
uv run bot.py                         # then open http://localhost:7860/client

Needs: Parakeet TDT 0.6B v3 (sherpa-onnx int8) in `~/.local/share/parakeet-dictation/models/desktop/`, Kokoro (`kokoro-v1.0.onnx`, `voices-v1.0.bin`) in `~/kokoro/`, and the `claude` CLI logged in. Paths are hardcoded at the top of `bot.py` for now.

Config (`config.toml`)

key	meaning
`provider`	`gemini`, `openai`, or `ollama` (front model)
`model`	model id, default `gemini-3.8-flash`
`base_url`	OpenAI-compatible endpoint (Ollama defaults to `localhost:11434`)
`api_key_env`	env var holding the key
`backend_cwd`, `backend_timeout`	where the backend agent works (default home) and its timeout (120 s)

Safety

The backend starts read-only (`Read/Glob/Grep` and a few `Bash(ls|du|find|stat|wc|cat)` commands). Anything else is blocked, the assistant tells you what it wants to do and asks yes/no, and only an explicit spoken "yes" (decided by code on your actual words, not by the LLM) re-runs it. Approving allows only the exact blocked command for that one retry. A "yes" only counts if you spoke after the question was asked. `find` is not allowed read-only because it can delete.

Using a subscription CLI this way depends on the vendor's terms; check them before relying on it.