Cursor adds two-way voice calls to coding agents Cursor adds two-way voice calls to coding agents

Speech and coding have different jobs

Cursor describes a division of labor inside the call. The voice agent handles the conversation and passes requests to the coding agent already attached to the chat. That coding agent retains its selected model. A spoken request therefore does not, by itself, replace the system producing the code. Progress reports can arrive while the conversation continues. The distinction matters when a team compares results: the speaking voice and the code-writing model are separate parts of the same interaction.

The handoff changes the input step

The documentation distinguishes a live call from dictation. Dictation produces editable text that the user reviews and sends. In voice chat, the conversational agent can hand off a request during the call. Calls require a signed-in Cursor account; an API-key session does not qualify. A transcript appears in the chat. Privacy Mode affects its live syncing across devices.

Privacy has a setting and a policy

The launch describes zero data retention when Privacy Mode is enabled. Cursor’s September 3 data-use policy explains the broader arrangement: customer data is excluded from its training, and it maintains retention agreements with model providers. Turning the mode off allows Cursor to store and use code-related inputs and actions for improving its models. The policy also describes temporary encrypted file caching to reduce repeated transfers. A voice interface adds another way to supply instructions, while the surrounding account settings still determine how customer data is handled.

The release follows hosted app publishing

A day earlier, Cursor introduced building and deploying live apps from cloud agents. That separate November 5 release can provision hosting, a Postgres database and authentication through Cursor sign-in. It covers new and existing repositories in Cursor Origin and GitHub. Publication requires an explicit user action after deterministic and agent-based security checks. Administrators can inspect apps and secrets, add review rules and remove an app from service. The two releases address different stages: a conversation supplies the request, while the deployment flow decides when an application becomes accessible to others.

A call is another way to supervise work

For a developer, the immediate change is the route between an instruction and the task already underway. An illustrative session might begin with a spoken repair request, continue with a question about the resulting change, and end with a written review of the diff. That sequence separates conversational convenience from checking the result. It also makes a practical comparison possible: the same small repair can be requested by voice or text, then judged against the same tests.