Your first conversation

Speak a real request#

Try the thing you would have typed:

Run the tests and tell me what broke.

Your words land in the agent as text. It gets to work, reading files, writing code, running tests, exactly as it would from a typed prompt. When it is done, it answers out loud in a natural voice generated on your machine:

Two failures in auth.test.ts, the new token refresh.

You never looked at the screen.

Interrupt it#

Full-duplex echo cancellation means you can talk over the agent while it is speaking and still be heard cleanly on open speakers. No push-to-talk gymnastics, no waiting for it to finish a paragraph you already understand.

SKI also detects when you have finished a thought, so it sends the sentence whole instead of chopping it at the first pause.

Add a screenshot to what you say#

Hit the screenshot hotkey and the capture rides along with your next spoken sentence. "Why is this button misaligned" becomes answerable, because the agent sees what you see. Your agent can also grab a screenshot itself when it needs one.

Keep it quiet when you need to#

Three controls live under a hover on the widget, and each can be bound to a hotkey:

• Mute cuts the mic at the source.
• Silent mode keeps the agent's replies as text only.
• Screenshots attach your screen to the next thing you say.

Next#

Make the controls yours in Hotkeys and quick controls. To understand what actually moved between the widget and the agent, read How SKI works.