SKI

SKI puts a voice on the coding agent you already use. Instead of typing into Claude Code, Cursor, Codex, Gemini CLI, Windsurf or OpenClaw, you speak — and instead of scrolling a wall of output, you hear the agent answer in a synthesised natural voice. The loop is full duplex with barge-in, so you can cut in mid-sentence and redirect it the way you would interrupt a colleague.

Everything that makes that loop work runs on your own machine. Speech recognition and neural voice synthesis are on-device, so no audio is uploaded and the app keeps working with no network. The same local pipeline powers the second half of the product: a meeting transcriber with speaker identification that turns calls into text without handing the recording to a transcription service.

The widget docks into the notch on macOS — including Macs that have no physical notch — and floats as a movable pill on Windows. Each project can be assigned its own voice, so you can tell which agent is talking without looking. A global hotkey starts and stops listening, screenshots can be captured on demand for the agent to look at, and an approve-before-send mode lets you review the transcript before it reaches the agent.

Requirements: Apple Silicon (M1 or later) on macOS 14.4 Sonoma or newer, or Windows 10/11 on x64. Linux is listed as coming.

Pricing: free for life on both platforms, no credit card. The one paid element is optional and separate — sending your agent into a Google Meet, Teams or Zoom call as a participant through AgentCall, billed per minute with free hours included.