Why We Built a Physical Keyboard for AI Coding Agents

A couple of years ago, AI image generators kept dreaming up a "keyboard for vibe coding." The renders went viral, everyone joked about it, and the meme refused to die. So we built the real thing.

Fifty days and three prototypes later, VibeKeys was shipping. Here is why we thought it was worth building — and what we learned.

The AI coding loop is decision-heavy, not typing-heavy. If you work with Claude Code or Codex daily, you already know the loop: describe what you want, let the agent work, review the result, then accept, reject, retry, compact, or go YOLO. Those decisions happen dozens of times a day, and today every one of them lives inside a terminal window.

Physical keys give those decisions muscle memory. VibeKeys puts accept, reject, retry, compact, YOLO, and push-to-talk on six dedicated keys, with a rotary knob for scrolling through generated diffs. Once the controls are always in the same place, the decision stops going through your conscious attention — the same reason camera shutters and game controllers are buttons, not menu items.

Hardware is the easy part; integration is the product. A keypad that sends keystrokes is a toy. So we open-sourced everything: the Rust firmware and the companion service (github.com/second-state), the CLI, and the integrations that hook the device into Claude Code, Cursor, Codex, Gemini CLI, OpenCode, and Copilot. Users remap keys, script behaviors, and extend it in ways we never planned.

Wireless changed how people use it. The Max console — an ergonomic key layout with hot-swappable Kailh Blue switches, a rotary knob, a built-in microphone, and a TFT status screen — doubles as a remote control: over WiFi or Bluetooth you can accept a change, reject it, or dictate the next instruction from another room while a long task runs. Several customers told us they now "supervise" agents from the couch, which honestly surprised us.

Where we are today: three models — Pro $29 wired, Plus $49 wireless, Max $99 console — free worldwide shipping, and an open roadmap driven by user feedback.

Question for the community: if you had six keys for your AI coding workflow, what would you map them to?