Open Whisperer adds a full voice loop to OpenAI's Codex CLI on your Mac: speak your prompt into the terminal, and hear each reply read aloud with streaming text-to-speech — at your chosen volume, hands-free, and entirely on-device.
# say "initiate", then just talk —
you ▸ add a --dry-run flag to the deploy script and explain what changed
…transcribed on-device, dictated into Codex CLI
codex ▸ Added --dry-run: it prints the planned steps and skips
the upload. I guarded it in main() before any writes.
▸ spoken aloud at your in-app volume, streaming
you ▸ perfect, commit it
On-device Whisper turns your speech into text and types it into Codex CLI in whichever terminal you use — Terminal, iTerm2, Warp, Alacritty, Ghostty, and more. Say "send" to submit.
The Codex hook speaks each response with Kokoro text-to-speech — streaming from the first sentence and honoring your in-app volume setting. "Hold on" stops it instantly.
Speech-to-text and text-to-speech run in-process on the Apple Neural Engine. Nothing about your voice or terminal session is sent to a third-party service.
MIT-licensed, no subscription, no account. The Codex CLI hook is right there in the source — audit it, tweak it, or wire it into your own workflow.
Download the DMG and drag Open Whisperer to Applications. On first launch it downloads the on-device Whisper and Kokoro models — fully native, no Python.
Use the one-click setup in the menubar to install the Codex hook. Pick your terminal as the auto-focus target so dictation always lands in the right window.
Hold Ctrl and speak, or say "initiate" for hands-free. Your prompt is dictated into Codex CLI and the reply is read back, streaming and at your volume.
Codex CLI lives in the terminal, which is exactly where reaching for the mouse breaks your flow. Voice keeps you in it: describe the change, listen to what Codex did, and confirm — without switching windows or scrolling back through output. For long agent runs, hearing the summary as it streams beats watching lines scroll past.
It shines for the iterative back-and-forth, too: correcting a plan, narrowing scope, or approving the next step is faster spoken than typed. And since transcription and speech both happen locally, the only network calls are Codex's own — Open Whisperer adds no cloud dependency of its own.
Also using Anthropic's assistant? The same app powers voice mode for Claude Code. Wondering how it stacks up against other dictation tools? See vs. superwhisper and vs. Wispr Flow.
Free, open source, 100% on-device. macOS on Apple Silicon.