What happened
Google’s Antigravity documentation says Voice Mode is a two-way spoken conversation powered by the Gemini Live model. It is available in the Antigravity App and Antigravity CLI; the CLI requires version 1.5.0 or later. A user can start it in a new or existing conversation, so an existing thread’s earlier messages remain part of the context. In the app, the user clicks Start Voice mode; in the CLI, they run /voice live or press F6. The CLI requires a microphone and speaker, with an additional setup path for SSH sessions using agy mic-serve and ANTIGRAVITY_MIC.
The interface is designed for ongoing work rather than dictation alone. A user can give the agent a task, correct or add details while it runs, ask for progress, stop a task, or interrupt speech. The agent works in the conversation, where written responses and work products remain visible. By default it works quietly and speaks when it has an answer or needs input; users can ask for spoken progress updates, which occur about every 30 seconds when enabled unless they request another interval. The app also offers voice-speed controls from 0.75x to 2x. CLI controls include voice speed from 0.5 to 3, volume from 0 to 100, and a full-duplex option for interruptions.
Voice can handle many approval interactions: the agent reads a permission request aloud, and the user can allow or deny it, including specifying a scope such as one action, one conversation, or one project. Google draws a firm boundary around destructive or elevated actions: commands such as rm, sudo, or force-pushing to a Git remote cannot be approved by voice, although they can be denied. Work involving files, commands, and tools remains subject to the conversation’s usual permissions.
The app streams microphone audio to Google servers, where Gemini Live recognizes speech and generates spoken replies. Its transcripts appear in a voice transcript card saved on the user’s computer and deleted when the conversation is deleted. Voice Mode is limited to local app conversations, cannot run in Remote Control or cloud sessions, and can be active in only one conversation at a time. In the CLI it is unavailable in print mode and remote-agent sessions. Google’s documentation says availability may not yet extend to every account.
What the archive adds
New Runtime’s earlier coverage on 2026-04-01 described Gemini Live as streaming audio directly to a model rather than using a speech-to-text, language-model, and text-to-speech chain. This is a new product integration of that voice capability in Antigravity, with an agent-workflow interface and explicit permission boundaries. Earlier coverage is related context, though the current documentation names Gemini Live without specifying a model version.
Trend signal
This feature is relevant to the assumption that real-time voice models can handle tool calls, interruptions, and session state within one model instead of relying on chained speech systems. Antigravity places spoken interaction around an agent’s existing conversation and work, but the documentation does not establish that the model itself performs tool calls: work and permissions remain in the agent conversation. The tracked assumption is therefore not confirmed by this release alone.
Why now
The product documentation presents voice as another way to interact with an agent during active work, including steering tasks and responding to permission prompts. It does not explain the rollout schedule or why the feature is launching now.
Why it matters
Voice changes the interaction cost of supervising a coding agent: a builder can describe a task, correct course, and ask for progress without returning to the keyboard. That may be useful when reviewing work or working hands-free, while the transcript and ordinary conversation preserve a textual trail. The permission design also clarifies that voice is an input channel, not a blanket authorization mechanism: destructive and elevated actions still require an on-screen decision. Teams evaluating voice interfaces should therefore assess both the convenience of live steering and the operational details around audio handling, session scope, and review. The feature’s local-session restriction and account availability mean it is not yet a universal remote-work path.
Limits and uncertainty
Google’s documentation does not establish how broadly Voice Mode is available across accounts, or quantify latency, recognition accuracy, or user outcomes. The documentation describes audio being streamed to Google servers and transcripts being saved locally, but it does not provide a retention or processing policy in the excerpt. The exact Gemini Live model version is also unspecified. Those details matter for teams assessing privacy, reliability, and rollout readiness.
Skill to build now
Try a bounded coding task in Voice Mode this week: ask the agent to find a failing test, instruct it to inspect before changing anything, then steer it toward a fix. Test a permission prompt and verify that a destructive action still requires on-screen approval. This gives a team a quick way to judge whether voice improves supervision without weakening its review process.
Reader takeaway
Antigravity Voice Mode makes speaking a practical control surface for an agent that is already working in a conversation. Treat it as a hands-free way to assign and steer work—not as a replacement for permission boundaries, written review, or checking whether your account and session type are supported.