Cursor adds two-way agent calls: voice handoff is not dictation—or read-only coding

Grok Voice handles conversation while the coding agent keeps its selected model. Understand call entry conditions, daily limits, mobile eligibility and permission boundaries.

AI IslandsPublished

Cursor's new call button does more than put spoken words into a draft message. Voice Chat with Agents, announced November 6, lets users assign work aloud, ask about progress and hear replies while a coding agent continues working. The important change is the handoff: the opportunity to review every word before sending is different, but reviewing the resulting code remains essential.

Conversation changes; the selected coding model stays

Cursor says Grok Voice handles listening and speaking, then hands coding work to the agent in the current chat. That coding agent keeps its selected model. Starting a call does not switch the entire coding task to Grok Voice or independently improve the coding model.

The feature works in the Agents Window, IDE, web and Cursor for iOS. Existing dictation still puts text into the input for the user to review and send; voice chat hands off work and provides spoken progress notes. Our analysis: dictation remains useful for exact paths, function names and constraints, while calls suit progress questions and coordination during a task. Neither replaces the other completely.

Check the entry conditions before changing plans

To start a call, wait for an idle agent and an empty input. An IDE chat also needs at least one message. Click the waveform labeled Start voice chat and allow microphone access. The adjacent microphone is dictation. The official Voice chat documentation provides troubleshooting details.

A signed-in Cursor account is required; API-key-authenticated sessions cannot start calls. The button is also hidden in subagents, Project workers, archived chats and Claude Code chats. Not being able to initiate a call from a worker chat is a different entry rule from switching Project workers during a call.

Calls have a daily limit, but this page publishes neither a numerical allowance nor a separate voice-plan or per-minute rate table. Do not assume unlimited free calls. The iOS app has its own eligibility: current documentation requires iOS or iPadOS 26.0 or later, lists supported paid or legacy plans, and specifies an English interface. Android remains planned. Desktop availability does not establish access on every phone or account.

Read-only questions do not make the whole call read-only

Cursor describes lookups of diffs, runs and agent status as read-only. The same voice interface can also hand modification requests to a coding agent. That read-only description belongs to information lookups, not all subsequent coding activity. File edits, tests, dependency installation and pushes still need the project's usual permission boundaries.

Start in a low-risk copy with a bounded request: “Explain which files this diff changes. Do not edit code or run commands yet.” Check the written record before approving the next action. For deletion, migration or deployment, put the scope in writing and review actual diffs and test results rather than accepting a spoken summary alone.

There is observable usage feedback: on November 7, a community user reported that voice handoff failed in a Project while a typed request started the work normally. This is a user report, not proof of a widespread defect or identical behavior across accounts. If a call produces no result, check whether the task actually reached the chat before retrying with typed text or reviewed dictation, avoiding duplicate assignments.

Read zero-retention wording with its setting and exceptions

Cursor's zero-data-retention statement for voice chat depends on Privacy Mode being enabled. Its general data-use page also describes exceptions for investigation after abuse detectors are triggered. With the mode off, some code and prompt data may be retained and used for improvement and training. “Voice is never stored” is therefore too broad.

Call transcripts appear in chat. Privacy Mode should not be read as automatically erasing every chat, cloud workspace or task record. Check organizational settings and permitted data flows before speaking client information, internal addresses or credentials. Our Cursor tool page provides the product entry point: the update changes communication, not authorization boundaries.