Inspiration Running a multi-million dollar e-commerce brand means drowning in operational tasks — reading support emails, tracking down orders, creating shipping labels, fulfilling orders across platforms. As COO of Gym Army, I spend hours every week on workflows that should take minutes. I wanted to build something that could just handle it — not through rigid automation, but through genuine AI understanding. Vox was born from that frustration. What it does Vox is an AI Voice COO that controls your Windows laptop through natural voice commands. You speak — Vox works. It reads your support emails in Spark, pulls up Shopify orders, creates Freightcom shipping labels, enters all the details, books the shipment, copies the tracking number, and marks the order as fulfilled in Shopify — all while narrating what it's doing in real time, asking for confirmation before any irreversible action, and maintaining full conversational context across your entire session. It's not a chatbot. It's not a script. It's an agent that sees your screen, thinks, and acts. How we built it Vox combines Anthropic's Claude computer-use API (claude-opus-4-5) with Playwright browser automation, giving Claude two sets of hands: pixel-level computer control for desktop apps, and fast DOM-level browser tools for web apps. Voice input runs through Groq's Whisper large-v3-turbo for fast, accurate transcription even in noisy environments. Responses are spoken back through ElevenLabs with a custom voice. The entire system maintains a persistent conversation thread across commands, so Vox remembers context the way a real COO would. Challenges we ran into Getting Claude to choose the right tool at the right time — browser automation vs. computer vision — required careful system prompt engineering. Whisper hallucinating in noisy hackathon environments forced us through multiple STT backends before landing on Groq. Chrome's remote debugging port refusing to bind consistently took significant debugging. And making the whole thing feel conversational — not robotic — required restructuring the agent loop to speak mid-execution rather than only at completion. Accomplishments that we're proud of A single voice command that triggers a fully autonomous, multi-platform workflow: Spark → Shopify → Freightcom → back to Shopify, with live narration the whole way. The system prompt is rich enough that Claude navigates complex multi-step forms — Freightcom's 4-step shipment flow, Shopify's fulfillment modal — with near-zero hesitation. And it sounds good doing it. What we learned Accurate voice recognition in noisy environments is a harder problem than it looks. System prompt design is as important as the underlying model — the difference between 30 iterations and 5 is mostly context quality. And real-world e-commerce operations have more edge cases than any tutorial prepares you for. What's next for Vox Full Gym Army operations integration: automated order triage, customer reply drafting, inventory alerts, and carrier rate optimization. Voice-controlled analytics. A persistent memory layer so Vox learns your preferences over time. And eventually — a version any operator can deploy for their own business.

Built With

  • claude
  • cursor
  • elevenlabs
  • groq
Share this project:

Updates