What is real-time-design?

Real Time Design is a terminal tool that lets a developer code by talking out loud instead of typing. A person opens it inside a project folder, speaks continuous coding instructions, and the tool turns that speech into structured tasks that get handed off to a coding assistant CLI, such as Codex or Claude Code, which then makes the actual edits to the repo. The voice side works through the OpenAI Realtime API, which streams microphone audio and uses voice activity detection to figure out when a spoken instruction is finished. It extracts a structured task from that speech, things like edit header, run tests, or create component, and if someone gives several instructions in one breath, it can split them into separate independent tasks. Up to four coding agents can run at the same time by default, so unrelated requests like changing a color, adjusting spacing, and running a typecheck can all happen in parallel. The voice model itself never writes any code, it only decides whether speech contains something actionable, and Real Time Design routes that decision to whichever coding CLI is doing the actual editing. A useful feature is barge-in correction. If someone says "make the header red" and then immediately says "actually make it blue instead," the first task is stopped and restarted using the correction. Saying something like "stop that" or "never mind" cancels a task outright instead of restarting it. The tool shows a compact terminal interface with the current listening state, the tasks that have been created, and each agent's status. Getting started requires Node.js 20 or newer, a working login for Codex or Claude CLI, an OpenAI API key for the realtime voice connection, a microphone, and on macOS the SoX audio tool if local capture needs it. After installing the package globally and adding an OpenAI key to a local .env file inside the target project, running the rtd command starts a voice coding session there. The author describes this as an early prototype. The core loop works, but barge-in detection is described as heuristic, agent progress is inferred rather than tracked precisely, and parallel agents can conflict if asked to edit the same files at once. Because it can edit real files, the README recommends committing or stashing work before using it heavily.