Voice mode
Talk to an agent pane over a real-time voice call, configure voice preferences, and review call history.
Voice mode opens a spoken conversation with a supported agent or worker pane. It uses OpenAI's Realtime API for audio while keeping the call attached to the pane you were already using.
Before the first call
Add an OpenAI API key under Settings -> API Keys. Then use Settings -> Call to choose:
- The realtime voice.
- Playback speed from 0.25x to 1.5x.
- The language used for the opening greeting.
Settings save automatically.
Starting a call
Click the Phone / Call button in a supported pane header. The overlay connects, loads the pane's latest transcript as context, and begins streaming captions and audio. If transcript loading fails, the call can still begin without that prior context.
Only one call can be active across AgentGrid at a time. Other call buttons remain unavailable until the current call ends.
Call availability follows the selected harness's capabilities. Claude, Codex, OpenCode, Cursor, Devin, and Grok support voice in the current harness catalog; Antigravity, Kimi, and Pi do not.
Controls
The overlay includes rolling captions, a mic-level waveform, mute, and hang up. You can copy individual transcript lines.
| Control | Action |
|---|---|
| CmdMCtrlM | Toggle mute while the call overlay is open. |
| EscEsc | Close the overlay and end the call. |
CmdMCtrlM is scoped to an open call; it does not start one.
Push-to-talk dictation
Push-to-talk is a separate in-app dictation shortcut. AgentGrid must be the focused application for it to fire.
- On macOS it is unbound by default.
- On Windows and Linux the default is bare AltAlt.
- Record any preferred single-key binding under Settings -> Shortcuts.
Push-to-talk is suppressed while a realtime call is active so two microphone paths do not compete.
Call history and handoff
AgentGrid stores completed call transcripts and summaries. Open Settings -> Call History to review calls by pane, read the transcript and summary, or reveal the underlying call log in the file manager.
When a call ends, AgentGrid generates a structured wrap-up. For a terminal-backed master session, that wrap-up is staged into the originating pane so the typed conversation can continue with the decisions and next actions from the call.
Related
- Coding harnesses - supported runtimes and capability differences.
- Mobile app - the mobile companion can connect to a paired desktop.
- Keyboard shortcuts - current default bindings.