SKI

Both halves of the conversation run on your computer: speech recognition listens, and a neural voice speaks. Nothing is uploaded because there's no cloud to send it to. The practical effect is that your audio and transcripts stay on your disk, and the local speech-to-text keeps the loop running offline. For anyone handling private code or client data, that removes a whole category of worry.

Dictation tools stop at text-in. SKI closes the loop by speaking the agent's reply back. Full-duplex echo cancellation means you can interrupt it mid-sentence on open speakers and still be heard, so you don't need push-to-talk or a headset. It figures out when you've finished and sends the thought whole, which makes back-and-forth feel natural. No grammar to learn. Just talk.

A green dot tells you the agent is actually listening, in real time rather than on a poll. You can bind several repos at once, even across different agents, and click the project name to switch. Each project answers in its own voice. In a session juggling three tasks, you can tell at a glance which one just responded.

Turn this on and every transcript lands in an editable bubble first. Nothing reaches your agent until you hit the checkmark, which gives you a chance to fix a misheard word or kill a bad request before it runs. It's the kind of guardrail that matters when an agent can write and delete files on its own. Worth flipping on early.

A hotkey grabs your screen, and the image rides along with your next spoken request. Your agent can also take a screenshot itself when it needs to see what you see, which matters because describing a layout bug or a visual glitch out loud is far slower than simply showing it. Why does that matter? Because a picture beats a paragraph every time.

SKI captures your mic plus the room audio and transcribes on-device, with speaker-tagged tracks and a live transcript while the call runs. Recording time and retention are unlimited, and you can export to Markdown or plain text. Because there's no bot joining the call, the other participants simply see you in the meeting, not a recorder.

Paste a meeting URL into the widget and your coding agent joins as a live participant: speaking, listening, presenting, or silently taking notes. It sees the screen, can screenshare, and reads and sends chat. Google Meet, Teams, and Zoom are supported, and you can spin up a meeting from the widget itself.

One skill covers the agents developers already run: Claude Code, Cursor, Codex, Gemini CLI, OpenClaw, and more. Connecting is one click per agent, so if you switch tools between projects you don't lose the voice loop. That breadth is what makes it a habit rather than a single-tool trick.