APP

ConvoBox

Local, backend-agnostic voice frontend for CLI coding agents

full-duplex voice control for Codex, Claude Code, OpenCode, and other agent CLIs

What it is

ConvoBox sits between you and whichever coding-agent CLI you are driving and lets you work by voice instead of, or alongside, the keyboard. The goal is a portable voice layer that can point at Codex, Claude Code, OpenCode, or another backend without being tied to one product.

It is local-first by default: speech-to-text and text-to-speech run on hardware you control, with a thin adapter layer translating spoken intent into each backend's native control surface.

Design goals

Full-duplex conversation

Continuous listening with voice-activity detection, barge-in, and spoken responses instead of push-to-talk dictation.

Backend-agnostic adapters

A small adapter surface maps text, interject, hard-stop, and busy-state operations onto each coding agent.

Local-first audio

VAD, STT, and TTS are designed to run on-device or on private-network hardware under your control.

Deterministic hard stops

Safeword handling stays out of the model path so an abort command cannot be second-guessed by an LLM.

Architecture

mic -> VAD -> local STT -> safeword check -> orchestrator -> backend adapter
                                                           -> local TTS

The orchestrator decides whether each utterance is a new command, a soft interjection into an active task, or a hard stop. Backend adapters prefer structured/headless interfaces where available and fall back to terminal control only when necessary.

Current implementation

Known gaps

Status

Scaffolding stage, with real pipeline validation underway. The project is not a stable end-user app yet, but the core audio/orchestration pieces are implemented and covered by automated tests.

See the GitHub repo for current implementation notes and testing status.