Cursor has added real-time, two-way voice calls with coding agents, using Grok Voice to handle listening and speaking while the coding agent continues its work. The 6 November changelog says Cursor voice chat works in the Agents Window, IDE, web and Cursor for iOS, allowing users to describe changes aloud and hear short progress updates.
This is different from dictation, which turns speech into text for the user to review and send. The new feature maintains a spoken conversation and hands tasks to the coding agent in the current chat. Cursor’s voice-chat documentation also sets an important access boundary: calls require a signed-in Cursor account, and sessions authenticated with an API key cannot start them.
Cursor voice chat separates conversation from coding work
Grok Voice supplies the speech-to-speech layer. The coding agent retains the model selected for that chat, so enabling a call does not mean every coding task switches to the voice model. During the conversation, the voice agent can pass a task or follow-up to the coding agent and continue speaking while that work runs.
The docs describe read-only lookups for questions about diffs, runs and agents. Calls can also search earlier messages and calls with the current agent, or switch to a different worker within a Project. That read-only description concerns those information lookups; it does not mean coding work handed to the coding agent is itself read-only.
The coding agent chooses short spoken progress notes and a brief completion answer rather than routinely reciting the full written response. If it provides no separate spoken update for a turn, the call reads the written reply. The transcript appears in the chat, keeping a text record alongside the conversation.
A waveform starts a call; the microphone still means dictation
Cursor’s documented call entry point is the waveform labelled Start voice chat, visible when the agent is idle and the chat input is empty. The adjacent microphone button starts dictation instead. In the IDE, an agent chat needs at least one message; the Agents Window also supports starting from a new chat. Microphone permission is required on the operating system, browser or phone.
On iOS, Siri can start a call with a new agent while the iPhone is unlocked. Live Activity controls offer mute and hang-up from the Lock Screen or expanded Dynamic Island. These are Cursor’s specified mobile call controls, not a claim that voice chat can control arbitrary applications on the phone.
The available named voices include Ara, Eve, Leo, Rex and Sal. Cursor also documents a daily voice-call limit without giving a numerical allowance on this page. By default, a call ends after about ten minutes of user silence when the agent has no work in progress.
Privacy wording belongs to the call feature and its setting
Cursor says that with Privacy Mode enabled, voice chat runs with zero data retention. Separately, its docs say cloud-agent transcripts sync to other devices during a live call unless Privacy Mode is on. Those statements describe the documented voice-chat behaviour; they are not an assurance about every category of data across the entire application.
For developers, the useful change is that conversation can continue while a task executes. Dictation collects a prompt before it is sent; a live call creates an ongoing channel for asking about progress and supplying follow-ups. The distinction can make it easier to discuss an evolving change, but it also makes the division between the voice agent and coding agent important. Spoken interaction is the interface, not a replacement for the underlying coding workflow or its selected model.
The account requirement is equally concrete. An API key can authenticate a session without making that session eligible to initiate this call feature. Likewise, read-only status questions and coding-task handoffs are different actions within the same conversation. Keeping those distinctions explicit avoids over-reading a voice interface as either a new all-purpose application controller or a different set of coding permissions. The privacy setting should be assessed at the same level of specificity: what Cursor says about calls under Privacy Mode, rather than a broader promise inferred from it.
The feature follows Cursor Live Apps’ cloud-agent building and hosting workflow. Grok Voice’s speech-to-speech release supplies the voice-model context, while Cursor’s Agent SDK concerns testable agent code projects rather than this spoken interface.
For users updating Cursor, the first distinction to check is therefore the waveform versus the dictation microphone—and whether the session is signed in with an account eligible to place the call.
Sources
Related stories
Cursor Live Apps: Cloud Agents Build and Host Web Apps gives the underlying agent-workflow context for a call that continues while coding runs.
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
