Typing a prompt to an AI coding agent is a lot like walking: controlled, deliberate, one careful step at a time. You form the whole thought in your head, then convert it into keystrokes one clause at a time, checking yourself along the way. It works. It gets you there. It's also not how the thought showed up in your head in the first place, you already had the sentence, whole, before your fingers caught up to it.
There's a faster way to move through a thought, and it doesn't involve typing it out at all. You just say it.
That's the whole pitch behind SKI, a free tool that gives Claude Code, Codex, Cursor, and other agents a voice in both directions, you talk instead of typing, and the agent answers you out loud instead of just printing text. Here's what actually setting it up looks like.
1. Download it
Grab it from https://heyski.io - Mac (Apple Silicon, macOS 14.4+) or Windows (x64). No Linux build yet. 2. Run onboarding
First run has you pick a widget, a floating pill, or a notch bar under the camera housing on a MacBook, choose a speech model, and install SKI's skill into your coding agents in one click. That skill install is what actually lets an agent understand voice input and know it can speak a reply.
3. Connect a project
Open your normal session, Claude Code, Codex, Cursor, whatever you're running, inside your project folder, and type ski once. The agent connects on its own; you'll see the project name and a green dot appear on the widget.
4. Talk instead of typing
Hold fn/Globe (or click the widget), say what you want, let go. SKI transcribes it locally and hands it to the agent exactly like typed text would be. No conversion step, no "let me phrase this properly", just the thought, as-is.
5. Let it answer instead of checking back
This is the actual point of the whole thing: the agent speaks up on its own when it's done or needs a decision, "tests pass, want me to commit?", and you answer by voice without walking back to the terminal. If you're running more than one project, each one can use a different agent and a different voice, so you know who's talking without looking.
6. Mute whenever
fn/Globe, Space after clicking the widget, a hotkey, or the widget's own mute button. It actually releases the mic at the OS level, the orange indicator on macOS turns off, which a lot of voice apps don't bother doing.
What's actually running where
Speech-to-text and text-to-speech both run on your machine. Nothing about your voice or your code gets uploaded, and it works offline. What you say to the agent still goes to that agent's own LLM, same as if you'd typed it, SKI is the voice channel, not the model.
The honest limits
It's built specifically for the agent loop (Claude Code, Codex, Cursor, Windsurf, Gemini CLI, Cline, Kilo Code, Continue, and about 16 more via a shared skill folder), not general dictation for arbitrary apps. English only for now. Apple Silicon only on Mac, no Intel builds.
The voice loop is free for life, no card, no caps. There's an optional paid add-on (AgentCall) if you want your agent to join a live Zoom/Meet/ Teams call as a participant or notetaker, recording a meeting locally and getting a transcript/summary is free and unlimited on its own.
Walking gets you there. This is the other option. If you try it, I'd genuinely like to know what breaks or what you'd want it to do next.