SKI Review (2026): Pricing, Features & Honest Verdict

SKI is a free voice interface for AI coding agents on Mac and Windows. Speech recognition and the synthetic voice both run on your own machine, so you talk to Claude Code or Codex and hear it answer back out loud. Best for: developers who already live inside a terminal coding agent. Price: Free forever, with a paid add-on for video meetings. Rating: 8/10.

What is SKI?

SKI is a small floating app that gives your coding agent ears and a voice. You speak, it converts your words to text locally, hands them to Claude Code, Cursor, Codex, Gemini CLI, Windsurf or OpenClaw, and then speaks the agent’s reply back to you in a neural voice. On a Mac it docks into the notch. On Windows it floats as a pill you can park anywhere. It launched on Product Hunt on 30 July 2026 and took the number one spot with 310 upvotes.

The company behind it is Pattern AI Labs, which also builds AgentCall, the meeting layer SKI uses when you want your agent to join a Zoom or Google Meet call. We found no named founder anywhere on the site, which is worth registering before you install anything that requests microphone access. What is unusual here is the honesty of the technical claim: SKI does not run a daemon or a cloud service. The widget and your agent communicate through plain files inside your own project directory, which means you can open them, read them, and delete them.

What Are SKI’s Key Features?

On-device speech recognition and on-device voice

Both halves of the loop run locally. Your audio becomes text on your machine, and the reply is synthesized on your machine with one of four local voices. The site claims 0 bytes uploaded, and the architecture supports the claim rather than just asserting it: there is no server component for the core product, so there is nothing for audio to be uploaded to. SKI works with no internet connection at all. Compare that to Wispr Flow, the category leader, which is cloud-first and sends your recordings to its own servers.

Full-duplex barge-in

This is the feature that separates SKI from dictation tools. Most voice tools are push-to-talk: you hold a key, you speak, they paste text into a box, and the interaction ends. SKI keeps the mic open while the agent is speaking and uses echo cancellation so you can talk over it on open speakers and still be heard cleanly. If your agent starts explaining something you already know, you interrupt it the way you would interrupt a colleague. It also does smart end-of-speech detection, so it waits for your full thought instead of firing off half a sentence.

A voice per project

You can bind several repositories at once and give each one a different voice. Ask a question and SKI routes it to whichever project you are addressing, and the answer comes back in that project’s voice. If you run three agents in three terminals across three repos, this solves the actual problem, which is not typing speed but knowing which agent just spoke. A green liveness dot shows whether the agent is genuinely listening, checked in real time rather than polled.

Approve-before-send and silent mode

Two controls that show the team has actually used this in a room with other people. Approve-before-send is an optional review mode where every transcript lands in an editable bubble first and nothing reaches your agent until you confirm it, which matters because voice recognition errors sent straight to an agent that can write files are a different class of problem than a typo in a chat box. Silent mode is the mirror of mute: mute stops the mic going in, silent stops the agent’s voice coming out, so replies appear as text in the pill while you sit in an open office or on a call. On Mac you can also hold the globe key as push-to-talk if you would rather not leave the mic open at all.

Meeting capture, both local and agentic

SKI records meetings two ways. The local recorder captures your mic and the system audio on separate tracks, transcribes on device, keeps recordings forever, and never puts a bot in the call. That part is free and unlimited. The second way sends your coding agent into the meeting as a live participant through AgentCall: it speaks, listens, screenshares, reads and sends chat, and can be told to sit silently as a notetaker instead. Connect Google Calendar and it joins a recurring series on time by itself.

Agent support is not uniform, and the site is upfront about it. Claude Code and Codex install at both user and project level. Cursor is project level only. Gemini CLI and OpenClaw are user level. Windsurf is manual, meaning you wire it up yourself. So the headline of six agents supported is accurate but the experience across them is not identical, and if Windsurf is your daily driver you should expect setup work the marketing does not describe.

How Much Does SKI Cost?

SKI is free forever on both Mac and Windows. You register once, no credit card, and every voice feature plus the local meeting transcriber is included with no usage cap. The only thing you pay for is sending your agent into a video call through AgentCall, which is billed by usage: the meeting bot starts at $0.35 per hour and real-time English transcription is $0.12 per hour, with six hours of meeting time included free and no subscription. You can also bring your own transcription or text-to-speech stack and pay only the bot rate.

That undercuts everything in the category by a wide margin. Wispr Flow is $15 per month or $144 per year. Superwhisper is $9.99 per month, and its lifetime licence went from $249 to $849 during 2026. SKI charges nothing for the equivalent capability and does more with it, because dictation tools do not speak back. The catch is not hidden but it is real: a free core product needs a business model, and here the business model is a meeting add-on that many developers will never touch.

Who is SKI Best For?

Use SKI if you run Claude Code, Codex or Cursor daily and you have started narrating your intent to the agent anyway. It is also a good fit if you work in a regulated environment where sending voice audio to a vendor cloud is a compliance conversation you would rather not have, because there is no cloud to send it to. And it is a good fit if you want to keep working while walking around, since the loop is genuinely hands-free in both directions.

Skip SKI if you want general-purpose dictation for email, documents and Slack. It is built for coding agents specifically, and Wispr Flow or Superwhisper will serve you better system-wide. Skip it if you are on an Intel Mac, because it needs Apple Silicon M1 or later on macOS 14.4 or newer. Skip it if you do not work in English, because version one supports English only. And skip it if you need a vendor with a published support commitment, because there is no named team here and no SLA.

Best SKI Alternatives

Wispr Flow is the mainstream option at $15 per month or $144 per year. It is polished, works everywhere on your system, and has real company behind it. But it is one-way dictation, it is cloud-first, and it will not read your agent’s answer back to you.

Superwhisper is $9.99 per month, or a lifetime licence that now costs $849. Like SKI it runs models locally on Apple Silicon with no audio leaving the machine, and it has years of polish behind it. It is Mac only, and again it stops at text-in with no spoken reply.

VoiceMode is free and open source, aimed squarely at Claude Code and other coding assistants. It is closer to SKI in intent than the dictation apps are. It also asks a lot more of you: you assemble the pieces yourself, and there is no notch widget, no project routing and no meeting capture.

Final Verdict: Is SKI Worth It?

Yes, and the cost of finding out is an installer and two minutes. SKI is the first tool we have tested that treats talking to a coding agent as a conversation rather than a dictation shortcut, and the two-way loop changes how the interaction feels. Running both directions on device is the right call, and it is not a marketing claim wrapped around a cloud service: the file-based architecture means you can verify it yourself. At $0 against $15 per month for the nearest mainstream competitor, the value question answers itself.

The reservations are about durability, not quality. Free forever is a promise from an unnamed team whose revenue depends on an optional meeting product, and the pricing page for that product publishes rates for the meeting bot and transcription but not for voice intelligence or text-to-speech, both of which it says are billed per minute without giving a number. The feature list has also wandered a long way from coding into calendar integration and meeting bots, which is usually a sign a team is hunting for the thing people will pay for. Install it, use the local half, and treat the meeting features as a bonus rather than a dependency. If Pattern AI Labs disappears tomorrow you lose an app, not your work.