A desktop app developed by US-based Pattern AI Labs. Instead of typing instructions to AI coding agents such as Claude Code or Codex, you speak them. What you say is converted to text on your machine and handed to the agent, and the agent’s reply is read back to you in a natural synthetic voice. Both speech recognition and speech synthesis run entirely on your own computer, so your voice data is never uploaded. SKI was released at the end of July 2026, and the version at the time of writing is v1.0.1. The app itself is free to use.
Key Features
- Instruct your agent just by speaking: Hold a hotkey (on Mac you can assign the globe/fn key for push-to-talk) and talk. Your speech is transcribed and delivered to the agent, which reads files, writes code, and runs tests, then answers you out loud
- All speech processing stays on the device: Speech recognition and synthesis run locally by design, so recordings are never sent to the cloud. It also works offline
- Full-duplex, so you can interrupt: You can speak while the agent is still talking — the back-and-forth feels like a phone call where you can cut in with a question
- Works with the major coding agents: Compatible with Claude Code, Cursor, Codex, Gemini CLI, Windsurf, and OpenClaw. You can connect several projects across different agents at the same time
- Per-project voice settings: Assign a different voice to each repository, so when you run work in parallel you can tell by ear which project just replied
- Meeting recording and screen capture: Supports local meeting transcription. You can also take a screenshot and send it to the agent together with your next spoken sentence
- Approve-before-send mode: An optional mode lets you review the transcribed text before it reaches the agent, preventing a misheard phrase from being executed as-is
Pricing
| Plan | Price | What’s included |
|---|---|---|
| App (Mac / Windows / Linux) | Free (for life) | All core features: voice input and output, agent integration, local transcription, screen capture |
| AgentCall (agent joins a video call) | Usage-based (free hours included) | Optional feature that lets the agent join video calls. Billed per minute; check the official site for the current rate |
Pricing is current as of August 2026. Please check the official site for the latest information.
Pros & Cons
✅ Pros
- Speech processing stays on the device, so what you say about work code never leaves your machine
- The core app is free, and you can start without signing up for a subscription
- You can issue instructions without taking your hands off what you are doing, easing strain on wrists and shoulders during long coding sessions
- Because the agent announces completion out loud, you notice it even when your eyes are on something else
- Not tied to a single editor — it works across multiple coding agents
⚠️ Cons
- English is the only supported language at present, so it is not suited to speaking Japanese
- macOS requires Apple Silicon (M1 or later) and version 14.4 or higher; it does not run on Intel Macs
- Not open source — it is proprietary software from Pattern AI Labs
- The benefits are hard to realize where you cannot speak aloud (shared offices, cafes, late nights at home)
- It is a young v1.0-era product; its track record and long-term support are still being established
Comparison with Similar Services
| Item | SKI | Wispr Flow | SuperWhisper | macOS Dictation |
|---|---|---|---|---|
| Main use | Voice conversation with coding agents | General dictation across any app | General dictation and transcription | OS-standard dictation |
| Agent reads replies aloud | Yes | No | No | No |
| Where processing runs | On device | Mostly cloud | Local models selectable | On device (when set to on-device) |
| Price | App is free | Free tier + paid plans | Free tier + one-time/paid plans | Free with the OS |
| Japanese | Not supported yet | Supported | Supported | Supported |
Who Is It For
- Developers who use Claude Code or Codex daily and want to swap typed instructions for spoken ones
- Anyone looking to cut typing volume and reduce strain on wrists and shoulders
- Teams handling code that cannot leave the company and who do not want voice data stored in the cloud
- People who want to look at other work while waiting for a long agent run to finish
- Developers comfortable giving instructions in English who want to try a new style of development
Summary
SKI is a desktop app for shifting your interaction with AI coding agents from typing to talking. Keeping speech processing on the device, and making the core app free, lowers the barrier to trying it considerably. The English-only limitation is significant, though: for Japanese speakers, whether you can comfortably phrase instructions in English decides how practical it is. It is worth trying once agent-driven development has become routine and the round trip to the keyboard starts to feel like friction.