Cursor Adds Two-Way Voice Calls With Its Coding Agents, With Listening and Speaking Handled by Grok Voice

Cursor now lets developers hold a spoken conversation with the coding agent working in their editor. The feature, Voice Chat with Agents, shipped on November 6 and runs in the Agents Window, the IDE, on the web and in the iOS app. It is a two-way call, not dictation, which earlier Cursor releases labeled Voice Mode.

The mechanics are deliberately narrow. When the agent is idle and the chat input is empty, the send button turns into a waveform labeled Start voice chat. The voice layer listens, replies and hands coding requests to the agent in that chat, while the coding agent keeps using whichever model the user already selected. On iOS, "Hey Siri, call Cursor" opens a call with a new agent, and mute or hang-up controls sit in the Lock Screen Live Activity and the Dynamic Island.

While the agent runs, the human can ask what changed in a diff, check on a run, search earlier messages and calls with that agent, or switch between workers in a Project. Those queries are read-only and do not interrupt the work in progress. The agent also narrates itself: short spoken progress notes during the task, and a brief answer when it finishes, rather than reading the full written reply aloud.

The voice stack is not Cursor's own. Listening and speaking run on Grok Voice, with a set of voices, Ara, Eve, Leo, Rex and Sal, selectable in Voice settings and switchable mid-call. With Privacy Mode on, Cursor says voice chat runs with zero data retention. Sessions authenticated only with an API key cannot start a call; a signed-in account is required.

There are caps. Calls carry a daily limit that Cursor surfaces when a user hits it, and by default a call hangs up after about ten minutes of silence if the agent has no work in progress, a guard against burning plan usage on an open line nobody is using.

The interesting part is not the microphone. Long-running agents have created a monitoring problem: the work takes minutes to hours, and the natural interface during that time is a conversation rather than a text box. Voice gives developers a way to redirect, interrupt or simply check status while their hands are elsewhere, the same bet Anthropic, OpenAI and Google are making with always-on agents that report back unprompted.

It also deepens a dependency worth noting. A coding tool used by professional teams now routes its real-time audio through a third-party model provider, so latency, availability and voice quality inside Cursor are partly xAI's problem. Cursor has not published latency figures for the feature, and the company has not said whether the voice layer will stay single-sourced.