Voice Agents

A voice agent is the same agent with the keyboard taken away.

Speech in, speech out, and in the middle a normal loop: decide, call a tool, speak. You are not talking to a nicer hold message. You are talking to something that can look up the order, move the booking, or — in this world — refuse to invent a stack and call MCP instead.

Talking is the easy part

Plenty of robots talk. Three harder things have to land at once or it feels haunted:

1. Latency. A half-second of dead air is a dropped call. It has to answer like a person, not a fax.
2. Turn-taking. You cut in. You change your mind. A good one goes quiet. A bad one talks over you.
3. Hands. Sounding warm is cheap. Doing the thing means tool calling into real software. No tools, no agent — just a mouth.

Use it when your hands are full. Do not use it to debug a type error. That still wants a keyboard, a CLI receipt, and a screen you can scroll.

This repo's plugin and npx create-hackhaton-space-stack@latest mcp are text-first hands. An Agents SDK is how some teams add ears and a mouth around that same loop. The builder stays a dining room. Nobody is choosing Hono by voice at 2 a.m. on purpose.

A voice agent can be confidently wrong out loud, with nothing on screen to audit. A good one knows the edge and hands you to a human.

What this unlocks

The door is no longer a keyboard. The employee is still the employee — tools, guardrails, a tight permission scope. You just talk.

When the hands are busy, the mouth has to be enough. When they aren't, type.