Show HN：SKI——面向 Claude Code、Codex 等的本地语音编程工具，免费使用

什么是 SKI？

这是一款适用于 Mac 或 PC 的小型悬浮应用，让你可以直接开口与 AI 编程助手交流，并以自然语音听到它的回答。

什么是“编程 Agent”？

编程 Agent 是开发者在代码项目旁运行的 AI 助手，例如 Claude Code、Cursor、Codex 或 Gemini CLI。它可以读取文件、编写代码、运行测试并回答问题。SKI 为它赋予了“耳朵”和“声音”。

SKI 实际如何工作？

你说话；SKI 会在你的设备上将语音转换为文本，并把它交给你的编程 Agent。Agent 随后完成工作——读取文件、编写代码、运行测试——再将回复发送回 SKI；SKI 会以自然语音朗读回复，并在药丸悬浮窗或刘海区中显示内容。这是在你自己的项目内、通过普通文件实现的一次完整语音往返交互。

什么是静音模式？

静音模式下，SKI 会继续监听并接收回复，但 Agent 的回答只会以文本形式显示在药丸悬浮窗或刘海区中，不会被朗读出来——适合安静的房间或通话场景。这与麦克风静音正好相反：麦克风静音会关闭输入端的语音；静音模式则关闭 Agent 输出端的语音。

我是否必须始终开启麦克风？

不需要——你可以随时将其静音，SKI 会停止监听并释放麦克风。在 Mac 上，你还可以使用 globe / fn 键：轻按可静音或取消静音，长按则可作为按住说话功能使用（仅限 Mac；需主动启用，并授予一项 macOS 权限）。

它与其他语音工具有什么不同？

大多数工具属于语音转文字听写——它们将你的话放进一个输入框，工作就结束了。SKI 则是为编程 Agent 设计的双向语音闭环：它会朗读回答，也会向你展示 Agent 正在执行什么操作。两者解决的问题不同。

我需要互联网连接吗？

不需要。监听、理解和语音合成都在你自己的设备上完成。唯一会连接互联网的功能，是让你的 Agent 加入视频会议——而这项功能是可选的。

我的语音会被录制或发送到其他地方吗？

不会。你的语音会在自己的设备上转换为文本，绝不会被上传。没有云端服务在监听——也没有任何内容需要上传。

价格是多少？

Mac 和 Windows 版本均永久免费——注册一次即可，无需信用卡。语音功能和会议转录功能均已包含在内。只有当你希望通过 AgentCall 让 Agent 加入视频通话时才需要付费；该功能按分钟计费，并包含免费时长。

SKI 如何录制会议？

有两种方式。本地录制——免费、无限制、在设备上完成：SKI 会将你的麦克风音频和会议音频捕获到独立音轨，并在你的设备上进行转录；不会有人加入通话，也不会上传任何内容（此模式不支持按说话人区分的转录）。通过 AgentCall 让你的 Agent 参会（云端）：一个机器人会以 Agent 参与者身份加入——可发言、回答、共享屏幕、查看屏幕、读取和发送聊天消息——或者作为静默记录员加入；它会生成按说话人区分的转录内容，并与你的项目连接。这是唯一的付费功能，按分钟计费，并包含免费时长。

我的转录内容和会话记录存储在哪里？

它们存储在内置的“Transcripts”窗口中，该窗口会列出你的会议、语音会话和本地录音。你可以在自己的设备上查看、编辑、复制、导出（Markdown 或纯文本）或删除这些内容。

运行它需要什么条件？

需要搭载 Apple Silicon 芯片（M1 或更新版本）、运行 macOS 14.4（Sonoma）或更高版本的 Mac，或者运行 Windows 11 或 Windows 10 的 PC。目前支持英语；更多语言正在规划中。

Mac 和 Windows 上是同一款应用吗？

是的——两端拥有相同的语音交互闭环、相同的 Agent 支持、相同的本地隐私机制，并且都可永久免费使用。在 Mac 上，它会停靠在刘海区中（即使是没有实体刘海的 Mac 也可使用）；在 Windows 上，药丸悬浮窗可停放在你指定的任意位置。Linux 版本也已列入计划。

What is SKI?

A small floating app for your Mac or PC that lets you talk out loud to your AI coding assistant — and hear it answer in a natural voice.

What's a "coding agent"?

An AI assistant developers run alongside their code — tools like Claude Code, Cursor, Codex, or Gemini CLI. It can read files, write code, run tests, and answer questions. SKI gives it ears and a voice.

How does SKI actually work?

You speak; SKI turns it into text on your machine and hands it to your coding agent. The agent does the work — reads files, writes code, runs tests — then sends its reply back to SKI, which speaks it aloud in a natural voice and shows it in the pill or notch. A full spoken round-trip, over plain files in your own project.

What is silent mode?

Silent mode keeps SKI listening and replying, but the agent's answers appear only as text in the pill or notch — not spoken aloud (handy in a quiet room or on a call). It's the mirror of mute: mute silences your mic going in; silent silences the agent's voice coming out.

Do I have to keep the mic on the whole time?

No — mute it any time and SKI stops listening and releases the mic. On Mac you can also use the globe / fn key: tap to mute or unmute, or hold it as push-to-talk (Mac only; opt-in, needs one macOS permission).

How is this different from other voice tools?

Most are speech-to-text dictation — they put your words in a box and stop. SKI is a two-way voice loop for coding agents: it answers out loud, and it shows you what your agent is doing. Different job.

Do I need an internet connection?

No. Listening, understanding, and speaking all happen on your own machine. The only feature that touches the internet is sending your agent into a video meeting — and that's optional.

Is my voice recorded or sent anywhere?

No. Your voice becomes text on your own machine and is never uploaded. There's no cloud service listening in — there is nothing to upload to.

How much does it cost?

Free forever on Mac and Windows — register once, no credit card. Voice features and the meeting transcriber are all included. You only pay if you want your agent to join video calls through AgentCall — per minute, with free hours included.

How does SKI record meetings?

Two ways. Record locally — free, unlimited, on-device: SKI captures your mic and the meeting audio on separate tracks and transcribes them on your machine; nobody joins the call and nothing is uploaded (no per-speaker diarization in this mode). Send your agent in via AgentCall (cloud): a bot joins as an agentic participant (talks, answers, screenshares, sees the screen, reads and sends chat) or a silent notetaker, with a speaker-diarized transcript, connected to your project — the one paid feature, billed per minute with free hours included.

Where do my transcripts and sessions live?

In a built-in Transcripts window that lists your meetings, your voice sessions, and your local recordings — view, edit, copy, export (Markdown or plain text), or delete, all on your machine.

What do I need to run it?

An Apple Silicon Mac (M1 or later) on macOS 14.4 (Sonoma) or newer, or a PC on Windows 11 or 10. English today; more languages planned.

Is it the same app on Mac and Windows?

Yes — same voice loop, same agents, same on-device privacy, free for life on both. On Mac it docks into the notch (even on Macs without a physical notch); on Windows the pill floats wherever you drop it. Linux is on the list.