Cursor lets you talk to its AI coding agent by voice, in real time

What's going on

Cursor is a popular AI code editor made by Anysphere. Since November 6, 2026, you can talk to its built-in coding agent. A coding agent is an AI that writes, edits and tests code on its own. The new feature is called "Voice Chat with Agents": a real-time conversation where both sides speak. You describe a change out loud, ask how the work is going, and the agent answers by voice too.

Voice chat is available in the Agents Window, the classic IDE (the development environment where you write code), the web version and Cursor for iOS. The key details in this article come from Cursor's changelog and documentation. Paul Lovelace summarized them on dev.to and openly says he hasn't tried the feature himself yet. So there's no independent hands-on test so far.

How voice chat works

There are actually two agents involved. A voice agent listens, replies and hands the programming task off to the coding agent in the chat. While the coding agent works, you can keep talking. You can ask what changed in a diff (the side-by-side comparison of old and new code), check on the status of a run, search earlier messages and conversations, or, in a project with multiple "workers," switch between these agents running in parallel.

The coding agent also speaks up. While it works, it gives short progress updates, and at the end it briefly sums up the result. It doesn't read out the full written response, which stays in the chat. Lovelace considers that the most important design decision:

"Voice is a terrible medium for long responses, so Cursor doesn't read you the full answer."

Under the hood, the voice part runs on Grok Voice, xAI's speech technology. The voice settings offer five voices: Ara, Eve, Leo, Rex and Sal. Grok Voice only handles speaking and listening. The coding is still done by the language model you've selected in Cursor. With Privacy Mode on, the docs say voice chat runs with "Zero Data Retention," meaning data isn't stored after processing. Lovelace recommends taking a close look at this before using it with client code.

How to start it, and what it isn't

Getting started takes a single button. When the agent is idle and the input field is empty, a waveform icon labeled "Start voice chat" replaces the send button. Click it to start the conversation.

On iPhone, you can go hands-free: "Hey Siri, call Cursor" starts the call. iOS app version 1.16.0 adds controls for the lock screen and Dynamic Island, so you can manage the conversation without unlocking your phone.

Don't confuse this with dictation, which used to be called "Voice Mode." Dictation just turns your speech into text in the input field. In the IDE, you toggle it with Cmd+Shift+Space. With voice chat, by contrast, the agent itself responds out loud.

The limits

Cursor is surprisingly upfront about the restrictions in its documentation:

• Account required: You need to be signed in to Cursor. If you only use Cursor with your own API key (an access key for an AI provider), you can't start voice chat at all.
• Daily limit: There's a maximum number of calls per day, but Cursor doesn't say what it is. Lovelace calls this a real gap, because you can't plan around it and heavy users only find out when they hit it.
• Auto hang-up: If the agent is idle and there's silence for about ten minutes, the call ends on its own.
• Button not everywhere: The start button is missing in subagents and project workers, archived chats, chats with Claude Code and teammates' chats.
• No new pricing: Cursor hasn't announced a separate voice plan or surcharge. The feature is simply part of Cursor as it stands.

Who it's for

Lovelace sees two main groups. The first is people who think out loud while coding. He compares it to "rubber duck debugging," where you explain your problem to a rubber duck to work out the solution yourself. Only this time the duck talks back and changes the code too. The second group is people who aren't at their keyboard, say, out and about with an iPhone while the agent works through a long task. According to Lovelace, the Siri integration and lock screen controls show that Cursor has exactly this scenario in mind.

If, on the other hand, you read every line of the agent's response and review every diff, you'll probably switch the feature off. Spoken summaries inevitably leave out details. And in an open-plan office, there's the question of whether you really want to talk to your editor out loud.

The Japanese platform note.com places the launch in a bigger trend: AI-assisted development is shifting from hand-writing code toward specifying requirements and reviewing results. But humans still need to check the quality and security of the finished code.

Background: Cursor is moving beyond the editor

Voice chat continues a trend that's been underway for a while. In May 2025, Cursor introduced "Background Agents" that handle tasks unsupervised. In June 2025 came a Slack integration and a web app that lets you control agents from the browser on desktop and mobile, as TechCrunch reported at the time. Cursor's head of product Andrew Milich told TechCrunch the goal was to reduce "friction" for users. Even back then, Anysphere reported more than $500 million in annual recurring revenue and customers at more than half of the Fortune 500.

The day before voice chat, on November 5, 2026, Cursor released "Build Live Apps." According to note.com, it lets agents build web apps and publish them complete with database, user authentication, hosting and a security review. The app only goes live once you explicitly click Publish. The feature is available on paid plans with cloud agents.

Our take

Two launches on two days show where Cursor is heading. The editor should not only write software but also ship it, and you should be able to talk to it instead of just typing. The fact that Cursor uses xAI's Grok Voice rather than its own solution shows how heavily the company relies on building blocks from other vendors.

A lot of important questions remain open, though. How good the voice quality is, how long you wait for replies, how Grok Voice copes with accents and technical jargon, and whether the spoken summaries actually help can't be gleaned from a changelog. The size of the daily limit is missing too. Whether talking while coding becomes a new habit or stays a gimmick you try once is something only real-world testing will show.