All docs

Voice control

Set up spoken or typed control of a browser or desktop app, review what Jev receives, and stop a session.

Ophio isn’t released yet. These docs describe the development version; app downloads are not available.

Voice control turns a spoken or typed command into clicks, typing and keystrokes on your computer, in a separate browser window or the focused desktop app. Jev, from TypeSafe, works out the steps using your own Jev access. This is different from dictating a task message or annotation.

Connect Jev

On the computer, run:

ophio jev connect
ophio jev status

Enter the key at the hidden prompt, never as a command argument or in task chat. The computer stores it locally; the phone sees connection state, not the key. Status checks the local configuration. Your first command tests whether the service accepts the key.

Jev has its own TypeSafe billing, separate from Claude Code or Codex subscriptions. Read the pricing and data-processing links in the voice control reference before connecting. An ordinary account has no zero-retention guarantee. Usage reports tokens, not a calculated price.

Turn it on

  1. Under Settings → Agents and accounts → Voice control, enable Turn on browser voice or Turn on desktop voice. These are separate switches and need management permission.
  2. For spoken commands, use ophio setup to install a local speech model. Settings → Voice reports speech availability and Language preference. Typed commands do not need speech recognition.
  3. In Control, use the microphone button labelled Control by voice and choose a target. An available Jev shortcut also opens voice control from full-screen desktop control.

Preview only · send no input is selected initially. Accept and preview commands starts without input. Clear that checkbox and review Accept and take over to let Jev act. Starting ends the current managed input on the target. Local keyboard and mouse input remain possible.

Use Hold to talk, or enter a Typed command and tap Send command. Read the Action log to check the result; a proposed action is not a completed one.

What Jev sees

Speech is transcribed on your computer. Jev receives the transcript and the names, roles and IDs of up to 100 on-screen controls. It does not receive screenshots, field values or general page text. Control names can themselves contain text from the page, and anything you say in the command is included.

Voice-control audio is not kept. The computer’s session record keeps the transcript, Jev’s answers, actions, results and usage. The separate Keep audio on computer setting for annotation dictation does not change this.

Safeguards

  • Destructive actions wait for confirmation.
  • Text, web addresses and key names must come from your command.
  • If a command could mean more than one control, you choose from a numbered list.
  • Voice typing refuses password and other sensitive fields. Use human control and the appropriate private step instead.
  • Each action checks the control’s identity and position again before acting.

Stop or disconnect

Stop voice control ends managed input. Check its confirmed state. Locking the computer, revoking the phone, losing the connection or ten minutes without a command also ends the session. Sessions have step and progress limits. An interrupted action is recorded as unknown and is not automatically retried.

To remove the stored key and disable voice control:

ophio jev disconnect

Disconnecting cannot retract commands already sent to TypeSafe.

Browser and desktop limits

Browser control uses a separate Chromium profile over local pipes, with no TCP debugging port. Ending voice control leaves its browser window open. Profile cleanup, including cookies and sign-ins, waits until the browser and its helpers exit; interrupted cleanup is recovered at the next start. Frames embedded from another site cannot be controlled yet.

Desktop voice control uses AT-SPI accessibility on Linux and UI Automation on Windows. The focused app must expose usable controls; there is no OCR fallback. If it cannot, use human desktop control. Implementation and automated tests are not complete physical-device qualification; see Platform qualification.