Voice
Talk to your agents out loud, dictate prompts, and have replies read to you.
Three ways to use your voice in Ghostex: the voice agent you talk with, Dictate to type by speaking, and Read aloud to hear replies.
The voice agent
Click the headset button in the header (left of Start), or press Option+Ctrl+V (Ctrl+Alt+Shift+V on Windows and Linux), to start a spoken conversation. Press it again to hang up. You can ask it to:
- Tell you what your sessions are doing, or which ones are waiting on you.
- Read you the last messages of a session. Name it the way you would say it (“the sidebar one”), or say “this one” for the session on screen.
- Send a prompt to a session. For anything risky, like deleting files, it reads the prompt back and waits for your yes.
- Draft a message without sending it: say “draft it”, then edit it and press Enter yourself, or say “send it”.
- Answer an agent’s question or approve a permission prompt.
- Start a new agent in a project you name.
Whatever session it talks about opens on screen, unless you’re typing. It tells you when a session finishes or starts waiting on you, and it sees an orchestrator’s threads as part of that orchestrator.
A small panel under the header shows live captions with Mute, Stop talking and Hang up. Drag it anywhere. Click the voice’s name to pick another voice, and the speed button (1× to 2×) to make it talk faster. It hangs up after 5 minutes of silence.
Memory: it remembers your last conversation as a short summary. A new call shows a Last conversation card: Continue picks up where you left off, Start fresh begins a new topic, and Forget deletes the summary.
Set it up
Open Settings > Voice agent, pick xAI Grok Voice (the default) or OpenAI GPT Realtime, and paste that provider’s API key. The key is checked, kept encrypted on your computer, and never shown again. The provider bills your account while a call is open.
Call mode is Regular (hands-free, you can talk over it) or Push-to-talk (it listens only while you hold Talk). On Grok, a regular call is billed for the whole time it’s connected, and push-to-talk only for the audio sent, so push-to-talk costs less.
It works without headphones and follows your computer’s default microphone. On the Android app, the headset button in the top bar talks to the voice agent on the computer you’re looking at, and keeps going with the phone locked.
Dictate
The microphone next to Send in the chat box (and on the terminal bar) turns your speech into text. Press it, talk, and press it again. The text goes into the chat box and is never sent on its own. Option+Ctrl+D (Ctrl+Alt+Shift+D on Windows and Linux) does the same.
Dictate uses xAI’s speech-to-text, so it needs an xAI key in Settings > Voice agent. It records first and transcribes after, so a failed transcription never loses what you said: Transcribe again, or rest the pointer on the microphone for Transcribe previous. Settings > Voice agent > Dictation history lists every recording to play, download, copy or transcribe again.
Read aloud
Read aloud speaks an agent’s final replies, so you can follow a session away from the screen. It’s experimental: turn on Enable Experimental Features (Settings > General > Advanced), then switch on Read aloud in Settings > Extensions > Store.
- Hover a reply and click Read aloud (the speaker).
- Turn on Auto-read in the chat’s ⋯ menu to hear every finished reply of the chat you’re looking at.
- Pick the voice and speed in Settings > General under Chat.
Code blocks and tables are skipped, and a reply that ends with a question says so. Read aloud speaks on the Mac and in a browser; the Windows and Linux apps can’t speak yet, but the phone app reads the chats of a computer that has it on.