Talking to your agents

Say it. Your agent gets it as text.

A prompt is slow to type on a phone keyboard. Hold the mic in Grenade and talk: your words become text on the phone, land in the field where you can fix them, and go to the agent when you send.

What you get

Hold, talk, fix, send.

Hold to talk, or tap

Hold the mic and speak until you let go. Or tap to start and tap again to stop, and talk hands-free.

You read it before it goes

What you said lands in the field, never straight in the agent. Fix a word, add a line, then send.

Pause to think

A long prompt keeps every word: what you said before a pause stays while you go on.

On your phone

Speech becomes text with the phone's own speech recognition, on the device where it supports that. The audio never goes to your Mac or to a relay.

Any agent

The text is typed into the session like any other: Claude Code, Codex or a shell.

On the Mac too

The Grenade Mac app has the same mic in each pane: click or hold, and the words land in that pane's field.
On the Mac

Nothing to set up

The mic is in the field of every session. The first time, the phone asks for the microphone and speech recognition. On the Mac, start the agent you want to talk to:

Not installed yet? Two steps on the Mac.

Mac
grenade new api --cwd ~/code/api --agent claude
Questions

Before you install.

Is my voice sent to a server?

Not to ours, to your Mac or to a relay. On phones that support it, speech becomes text on the phone. On those that do not, the phone's maker processes the audio (on an iPhone, Apple). The privacy policy has the details.

Does the agent hear me?

No. The agent gets text, the same as if you had typed it.

Can I answer a plan by voice?

Yes. The field under a plan has the mic: say what should change and send it back.

Is the mic on while I am not using it?

No. It is used only while you tap or hold the mic button.

Related

Where to go from here.