SKI

SKI is a desktop voice layer for AI coding agents rather than a coding model of its own. It lets a developer speak instructions to a connected agent and hear the agent answer aloud, with support advertised for Claude Code, Cursor, Codex, Gemini CLI, OpenClaw, Windsurf and additional skill-compatible agents. The core speech-to-text and text-to-speech loop is described by the vendor as running on-device, and the agent connection uses a skill plus ordinary files inside the project rather than simulated keyboard input.¹²

The product goes beyond simple dictation. Current features include multiple projects and agents, global hotkeys, an optional approve-before-send step, full-duplex interruption/echo cancellation, agent-liveness cues and local meeting recording/transcription. A separate AgentCall feature can send an agent into Zoom, Meet or Teams as a participant; that cloud workflow should be treated as a distinct product surface because it does not share the fully local architecture of the ordinary voice loop.¹

SKI says its local voice workflow and local meeting transcriber are free forever, with one-time registration and no card required. The only paid surface currently described is AgentCall: cloud meeting participation is billed by the minute after included free usage is exhausted.¹³ Any subscription or API charges for the connected coding agent—such as Codex, Claude Code or another provider—remain separate from SKI.²

The current download surface covers macOS 14.4+ on Apple Silicon, Windows 10/11 x64 and Ubuntu/Debian x64 Linux. Linux is the newest platform: the Product Hunt maker forum announced the Linux release in August 2026 and explicitly described it as early, with rough edges still expected.¹³

The strongest evidence is first-party. SKI states that local speech recognition, spoken replies and local meeting recordings stay on the device and that the normal voice loop works without an internet connection.¹ The official GitHub repository adds useful distribution detail: it hosts signed release artifacts and update metadata, but explicitly says it contains binaries only and not the application's source code. SKI should therefore not be described as open source.⁴

External validation remains shallow. The Product Hunt launch reached the number-one daily rank and attracted roughly 1,400 followers, but launch attention is not evidence of transcription accuracy, agent reliability or productivity gains.³ Product Watch independently confirms the product identity and launch-era proposition but does not provide substantive comparative testing.⁵ Current independent coverage does not provide an exact-product benchmark or verified-user rating that can safely be transferred to SKI's desktop voice surface.

The clearest fit is a developer who already uses a coding agent and wants voice to become a two-way interaction channel rather than a one-direction dictation tool. It can be useful for hands-free prompting, accessibility, switching among active coding projects or capturing a meeting locally and then asking the coding agent to work from the transcript.¹

It is a weaker fit for developers who prefer keyboard-first workflows, work in shared/noisy environments without a headset, or require a mature independent evidence base before adopting a new always-listening desktop utility. Teams considering AgentCall also need to evaluate that cloud service separately from the local SKI voice loop because its infrastructure, pricing and data path differ.¹³