Cursor adds real-time voice chat with agents on desktop, web and iOS

2 hours ago

8 min read

Cursor has added real-time voice chat with its coding agents, letting developers hold a spoken, two-way conversation with an agent while it works instead of typing every instruction into the chat box. The feature was published in Cursor's official changelog on November 6, 2026, under the title "Voice Chat with Agents", and it is available in the Agents Window, in the Cursor IDE, on the web and in Cursor for iOS once the app is updated.

The listening and speaking side of each call runs on Grok Voice, the speech-to-speech technology from xAI, while the coding agent itself keeps whatever model the user has selected for it. During a call a developer can describe a change out loud, ask what an agent is doing, review what changed in a diff, check the status of runs and hear progress updates as the work moves forward. When the agent finishes, it gives a short spoken answer, and the full transcript of the call is written into the chat.

Cursor also states that with Privacy Mode turned on, voice chat runs with zero data retention. On iPhone, a call to a new agent can be started hands-free by saying "Hey Siri, call Cursor", and an active call can be muted or ended from the Lock Screen. Voice calls are subject to a daily cap, and a call ends automatically after about 10 minutes of silence from the user when the agent has no work in progress.

··········

HOW CURSOR VOICE CHAT WITH AGENTS WORKS.

Voice chat is built as a layer on top of an existing agent conversation. The voice agent that answers the call works alongside the coding agent. It listens to the developer, hands coding tasks or follow-up requests to the coding agent in the current chat, and reports back what that agent produces. According to Cursor's documentation, the voice agent can answer read-only questions about diffs, runs and agents, search earlier messages and earlier calls with the same agent, switch workers inside a Project that has workers, and hang up when the user says the conversation is over.

The coding agent decides what is spoken aloud. In practice that means short progress notes while it works, a brief answer when a task is complete and updates from any workers it started. If the coding agent writes no spoken update for a turn, the call reads out its written reply instead, so the developer is never left without feedback. Everything said during the call ends up in the chat transcript, which keeps a written record of the instructions given by voice next to the code changes they produced.

Starting a call follows a fixed sequence. The user opens an agent chat (in the Agents Window any chat works, including a brand-new one, while in the IDE the chat needs at least one message), clears the chat input and waits for the agent to be idle. A waveform button labeled "Start voice chat" then appears where the Send button normally sits; while the agent is running, the Stop button occupies that position instead. After the user grants microphone access, the call connects, and if the user stays silent the agent checks in after a few seconds.

Voices are configurable. Cursor names Ara, Eve, Leo, Rex and Sal among "many named voices", and a default can be set in the Voice or Agent Voice settings. The voice can also be changed during a call, and the choice is stored for that agent, so an agent with its own saved voice keeps using it in later calls.

··········

CURSOR VOICE CHAT VS DICTATION: WHAT EACH ONE DOES.

Cursor already offered speech input before this release, and the changelog draws a clear line between the two features. Dictation, which earlier Cursor releases called Voice Mode, turns speech into text in the chat input so the developer can review it before sending. Voice chat is a conversation: the voice agent sends the prompt on the user's behalf and answers out loud.

........

Feature	Voice chat	Dictation
What you get	Spoken two-way conversation	Text in the chat input
Who sends the prompt	The voice agent	You, after review
How to start	"Start voice chat" waveform button	Microphone button
Keyboard shortcut	None	Cmd+Shift+Space in the IDE, hold Ctrl+M in the Agents Window
iOS	"Hey Siri, call Cursor"	Record voice
Spoken replies	Yes	No

........

The distinction has practical consequences. Dictation suits precise prompts where wording matters, such as a long refactoring instruction with file names and constraints, because the text can be corrected before it reaches the agent. Voice chat suits supervision: asking an agent what it has changed, redirecting it after a test failure or checking on several workers while away from the keyboard, where speed of interaction counts for more than exact phrasing.

··········

CURSOR VOICE CHAT AVAILABILITY, REQUIREMENTS AND LIMITS.

The feature works on four surfaces: the Agents Window, the IDE, the web app and Cursor for iOS. Cursor's instruction is simply to update the app to try it. The changelog entry and the documentation do not list specific plans, a separate price or a per-minute charge for voice chat in agent conversations, and they do not give a number for the daily call cap. When a user reaches that cap, Cursor says it shows when the next call can be made.

........

Item	What Cursor documents
Surfaces	Agents Window, IDE, web, Cursor for iOS
Account	Signed-in Cursor account; API-key sessions cannot start voice chat
Microphone	Permission required in system or browser settings
Speech technology	Grok Voice; the coding agent keeps its selected model
Silence timeout	About 10 minutes with no agent work in progress
Daily limit	Yes, number not published
Privacy Mode	Zero data retention for voice chat
Pricing	Not stated for agent voice chat

........

There are also places where the voice button does not appear. Cursor lists subagents and Project workers, archived chats, Claude Code chats and chats owned by another teammate. In the IDE only, the button is also hidden for cloud agents that are still streaming and when dictation is turned off. If none of these conditions applies and the button is still missing, Cursor's advice is to update the app.

For developers comparing what they pay across tools, our guide to AI coding agent pricing for Cursor, Claude Code, Codex, GitHub Copilot and Grok sets out the plan structures. Cursor's current pricing page lists Free at $0, Plus at $8 a month, Premium at $30, Super at $100 and Ultra at $200, with Teams seats at $30, $100 or $200 per month. Its FAQ says Grok Chat, Imagine, Voice, Grok Bot and Cursor draw from the same weekly usage pool for Cursor models, with usage rates that vary by product, although the voice chat documentation itself does not describe how agent calls are metered.

··········

CURSOR VOICE CHAT ON IOS: SIRI, LOCK SCREEN AND DYNAMIC ISLAND.

The iPhone version adds controls that do not exist on the desktop. Saying "Hey Siri, call Cursor" starts a call with a new agent, provided the iPhone is not locked. When iOS cannot show the call on the system call screen, it appears on the Lock Screen and in the Dynamic Island, and the user can mute or hang up from the Lock Screen or from the expanded Dynamic Island. The compact Dynamic Island does not show those controls.

This makes the phone a supervision device for agents running elsewhere. A developer who has started a long cloud agent task at a desk can check on it later by voice, ask for a summary of the diff and hand over a follow-up instruction without opening a laptop. Cursor notes that during a live call, cloud agent transcripts sync to the user's other devices, unless Privacy Mode is on, so the conversation remains available as text on the desktop afterward.

··········

CURSOR VOICE CHAT PRIVACY MODE AND DATA RETENTION.

The privacy statement in the changelog is short: with Privacy Mode on, voice chat runs with zero data retention. Cursor's general data-use page explains what that mode means across the product. With Privacy Mode enabled, Cursor says customer data is not used for training by Cursor, model providers agree not to train on it, and Cursor has zero data retention agreements with its providers, with an exception for prompts flagged by abuse-detection classifiers, which may be kept for investigation and then deleted under each provider's policy. With Privacy Mode off, Cursor may store and use prompts, code snippets and editor actions to improve its features and train its models.

For teams the relevant point is that spoken instructions are treated as part of the same data flow as typed prompts. A voice call that mentions customer names, internal systems or credentials ends up in the transcript, so organizations that already require Privacy Mode for code should apply the same policy before allowing voice calls. The documentation does not say how long voice transcripts are stored when Privacy Mode is off.

··········

WHAT GROK VOICE COSTS DEVELOPERS OUTSIDE CURSOR.

Cursor does not publish a per-minute rate for voice calls with its agents, but the underlying technology has a public API price for developers who want to build similar features into their own applications, for example with Cursor's Grok Voice plugin, which adds real-time speech-to-speech, dictation and read-aloud skills to an app. In its October 14, 2026 announcement of Grok Voice 3.0, xAI listed API pricing of $0.08 per minute, or $4.80 per hour, and claimed time to first audio under one second. Cursor's changelog does not say which version of Grok Voice powers agent calls.

At that API rate, a voice assistant used for a 15-minute call per working day would cost $1.20 a day, or $26.40 over 22 working days, per user. A team of 10 developers each averaging 30 minutes of calls a day would spend $24 a day, or $528 over the same 22-day month. These figures apply only to a self-built integration on the xAI API; inside Cursor, voice chat follows Cursor's own plans and usage limits, which the documentation does not quantify for agent calls.

The new feature also connects to other recent changes in Cursor's cloud agents, including the ability to build and host live web apps with Postgres and sign-in, announced one day earlier. Together they move more of the developer's interaction with agents away from the editor window and toward supervision and review.

··········

CURSOR VOICE CHAT WITH AGENTS: CONFIRMED FACTS AND OPEN QUESTIONS.

Cursor has confirmed the essentials: real-time voice calls with agents on the Agents Window, IDE, web and iOS; Grok Voice for listening and speaking while the coding agent keeps its own model; read-only questions about diffs, runs and agents; search across earlier messages and calls; worker switching inside Projects; Siri activation and Lock Screen controls on iPhone; a silence timeout of about 10 minutes; and zero data retention when Privacy Mode is on.

Several points remain undocumented. Cursor has not published the size of the daily call cap, has not said whether voice chat is limited to particular plans or how calls count against usage, and has not stated which Grok Voice version or which languages are supported in agent calls. Voice chat is also unavailable in Claude Code chats, subagents and Project workers, which limits it to the main agent conversation for now. Developers who rely on precise, reviewable prompts still have dictation, while voice chat gives them a faster way to monitor and redirect agents that are already running.

··········

·····

DATA STUDIOS

·····

datastudios.org