Coding agents flipped the bottleneck. The model writes faster than you can read; the slow part is you, typing the paragraph that explains what you actually want. Most people type around 40 wpm and speak around 150. That gap is the whole pitch.
Vokk is push-to-talk dictation for macOS. Hold a key — fn on laptops, ⌃⌥ on desktops, customizable — talk, release. The words are typed into whatever has focus. Codex's prompt box counts. So does everything else.
Codex CLI ships a dictation affordance in its prompt box, and it's a genuinely nice touch — one less thing to install, right where you compose the prompt. If that's all you ever need, you don't need Vokk.
The difference is scope. A built-in only works inside the box it's built into. Vokk is one key that works everywhere:
rg patternVokk types locally into whatever window has focus, so it works in any terminal window regardless of what that terminal is connected to.
And because a coding agent reads intent rather than syntax, dictation is unusually forgiving here. A stray "um", a self-correction, a comma in the wrong place — Codex handles it. You're describing a change, not writing a formal spec.
Vokk types what you said. No LLM rewrites your words on the way to the prompt, which matters when half of what you're saying is a file path or a symbol name.
Transcription runs on your Mac — whisper.cpp on the GPU via Metal, well under a second. No sign-up, no account, no telemetry. Your prompts about your employer's private repo don't leave the machine.
The airplane test: turn Wi-Fi off, hold the key, it types anyway. Optional cloud transcription (Groq) exists behind a license, but local is the default.
codex, hold the key, talk.