Vocl

Voice-native interface for Claude Code -- talk to your codebase.

Quick Start

npx vocl

On first run, you'll be prompted for your LiveKit Cloud credentials. That's it -- the browser opens automatically and you're talking to Claude Code.

What You Need

• Node.js 18+
• LiveKit Cloud account -- cloud.livekit.io (free tier available)
• Claude CLI installed and authenticated (claude on your PATH)

How It Works

You (browser mic/speaker)
  -> LiveKit Cloud (WebRTC + Inference Gateway)
  -> Vocl agent (Node.js)
      - STT: Deepgram Nova-3
      - LLM: openai/gpt-4.1-mini (Layer 1 -- conversation + routing)
      - Tool: claude -p subprocess (Layer 2 -- engineering work)
      - TTS: Deepgram Aura-2
  -> You (browser speaker)

One process runs everything: the voice agent, the token server, and the static UI. All STT/LLM/TTS go through LiveKit's Inference Gateway -- one API key, no separate provider accounts.

Configuration

Config is resolved in priority order: env vars > .env file > ~/.vocl/config.json

The first-run setup wizard saves to ~/.vocl/config.json. You can also set environment variables directly:

export LIVEKIT_URL=wss://your-project.livekit.cloud
export LIVEKIT_API_KEY=APIxxxxx
export LIVEKIT_API_SECRET=xxxxx

CLI Flags

npx vocl                          # start everything, open browser
npx vocl --no-open                # start without opening browser
npx vocl --project-dir ~/myrepo   # set Claude Code working directory

VOCAL_CLAUDE_PROJECT_DIR env var also sets the working directory (defaults to cwd).

Examples

• "Hey, what's up?" -- answers directly (no Claude Code)
• "What files are in this project?" -- calls Claude Code, speaks back a summary
• "Dispatch a security review" -- Claude Code dispatches background task
• "How's that review going?" -- asks Claude Code for status