SKI

SKI is a floating desktop widget for Mac and Windows that turns your AI coding agent into a two-way voice conversation partner. Built by Pattern AI Labs, it runs entirely on your own machine — speech recognition and neural voice synthesis both happen locally, so nothing you say is ever uploaded. Version 1.0.1 launched in late July 2026 and is free for life with a one-time registration.

What It Is

SKI is a voice coding interface — not a dictation tool. Where most voice tools stop at converting speech to text, SKI closes the full loop: your spoken request goes to your coding agent, the agent does the work, and its reply is spoken back to you in a natural neural voice. The app sits in the Mac notch (or draws its own on Macs without one) and floats as a draggable pill on Windows, staying out of the way until there's something to say.

How the Voice Loop Works

The workflow is designed to go from install to conversation in under two minutes:

• Install — a DMG on Mac or EXE on Windows; no credit card required.
• Connect your agent — one click per agent using a small "SKI skill" your agent already understands.
• Talk — the floating pill shows it heard you, the agent gets to work, and the reply is spoken aloud.

Full-duplex echo cancellation means you can interrupt the agent mid-sentence and still be heard cleanly on open speakers — no push-to-talk required. An "approve-before-send" mode lets you review every transcript in an editable bubble before it reaches the agent.

Privacy Architecture

SKI is built around a strict on-device model:

• Speech-to-text and neural voice synthesis run entirely on your machine.
• The widget and agent communicate through plain files inside your own project — no daemon, no cloud service.
• Audio is never uploaded; there is nothing to upload it to.
• Transcripts, sessions, and local recordings all live in a built-in Transcripts window on your machine, exportable to Markdown or plain text.

Agent and Platform Support

SKI supports six coding agents at launch: Claude Code, Codex, Cursor, Gemini CLI, Windsurf, and OpenClaw. Each project can be bound to its own voice, so multiple repos can be active simultaneously and each answers in a distinct voice. The app also supports screenshots on demand — a hotkey grabs your screen and attaches it to your next spoken request, and the agent can request a screenshot itself.

Meeting Features

Beyond coding sessions, SKI includes two meeting modes. The local meeting transcriber captures mic and system audio from Zoom, Meet, or Teams entirely on-device — unlimited recording time, zero cloud, with live transcription and export. The second mode, powered by AgentCall, sends your coding agent into a video call as a live participant: it can speak, listen, screenshare, read and send chat, and produce a speaker-diarized transcript. The AgentCall integration is the only feature that touches the internet and is the only paid component.

Update: v1.0.1

The current release is SKI 1.0.1, published July 27, 2026, available as a notarized DMG for Apple Silicon Macs (macOS 14.4 Sonoma or later) and an x64 installer for Windows 10/11. The GitHub releases repository (pattern-ai-labs/ski-releases) hosts binaries and the auto-update feed; the app checks for new releases automatically and verifies each update cryptographically before applying it. Linux support is listed as planned.