Cursor and Google Give Coding Agents a Voice, Claude Code Stays Silent

On Friday, two of the biggest players in AI‑assisted development shipped the same feature on the same day: two‑way voice for coding agents. Cursor launched Voice Chat with Agents, powered by xAI’s Grok Voice. Hours later, Google announced Voice mode in Antigravity, running on Gemini Live. Both let developers speak to their coding agents and hear them talk back — a shift from the type‑prompt‑wait‑read loop that defines most agent interactions today.

The timing follows OpenAI’s move earlier this year. Codex added two‑way voice to its desktop app in July and enabled live voice conversations in the CLI by default in September, The New Stack reported. Anthropic’s Claude Code, by contrast, still only offers one‑way dictation — the /voice command turns speech into prompt text, and the agent replies in writing.

The key innovation is that the voice channel stays open while the agent works. In Google’s Antigravity, a developer clicks a waveform button to start a call, tells the agent what to do, and the agent replies aloud. The call remains active throughout the task — the developer can ask for status updates, interrupt mid‑sentence to change direction, or approve commands without touching the keyboard.

“Some of the best engineering happens out loud,” wrote Nick Van der Auwermeulen, developer relations engineer at Google, in the Antigravity blog post. “You talk a bug through with a teammate, think out loud in front of a whiteboard, or say ‘wait, try it the other way’ halfway through a pairing session.”

Cursor’s implementation makes a similar pitch: “Describe a change out loud, ask what an agent is working on, and hear updates as the work moves forward,” the company wrote in its changelog. One difference: Cursor supports voice on mobile (iOS) and web, while Google offers it in the desktop app and CLI.

The voice features tap into different underlying models. Cursor uses Grok Voice, which xAI launched through its API in April. Google uses Gemini Live, its real‑time conversational AI. OpenAI’s Codex voice runs on its own voice stack, integrated into the ChatGPT desktop app.

Cursor’s integration is notable partly because of the company’s trajectory. Acquired by SpaceX for $60 billion in August, Cursor has been iterating rapidly — it added dictation in October 2025 and updated voice calls last month to let agents give short spoken progress updates, according to Cursor’s release notes. The move to full two‑way voice signals that Cursor sees voice not as a novelty but as a core interaction model for coding agents.

Anthropic’s decision to keep Claude Code voice as dictation‑only stands out against this backdrop. The Claude app has two‑way voice, but Claude Code specifically does not. The /voice command converts spoken words into prompt text, and the agent replies with written output. There’s no spoken response, no live call, no mid‑task interruption.

Whether this is a product philosophy choice or a temporary gap is unclear. Anthropic has been busy shipping frontier models — Claude Fable 5.2 and Mythos 5.2 launched on Monday — and the company may simply be prioritizing model capability over interaction polish. But the gap is real: developers using Claude Code can dictate, but they can’t have a conversation.

The move to voice changes the ergonomics of working with coding agents. Instead of the type‑prompt‑read loop, developers can talk through a problem while the agent works, checking in verbally without context‑switching to a chat window. For long‑running agent tasks — refactors, multi‑file changes, debugging sessions — the ability to hear progress updates and redirect mid‑stream could reduce the friction that makes agents feel like a separate tool rather than a collaborator.

The race is now between three platforms (Cursor, Antigravity, Codex) offering conversational coding agents, with Claude Code the notable holdout. For developers who prefer voice interaction, the choice of coding tool may increasingly come down to which agent can hold a conversation — not just which one writes the best code.