I built a voice coding agent to skip the typing

I built Rubber Duck because typing was the slow part. It streams audio to OpenAI Realtime with gpt-4o-mini, runs tool calls, and drops what I ramble into Things 3. Then Commandment, an open source WisprFlow alternative, in one night. Same annoyance both times: when the thing listening is an LLM, the mouth beats the fingers.

Two voice tools

You feel it after a week. Talking all day is uncomfortable in its own way, but I keep reaching for the mic instead of the keyboard. That’s the tell.

The translation layer

You don’t work out how to type what you mean. You say it, and the agent does it. The friction was never words per minute. It was turning a thought into structured input. Voice skips that step.

Until the tooling catches up, a high performing keyboard is still the fastest physical interface between thought and shipped code.

What’s left

Every tool I build now points the same way. Speech will replace typing for AI work. The open question is how fast the mic catches up to the demand, which is already here.