Cursor adds voice chat for coding agents, with daily limits
Cursor users can now speak with a coding agent, give it work and hear progress updates while it runs. Announced on November 6, the new voice chat works in the IDE, Agents Window, browser and iOS app. Grok Voice powers the spoken conversation.
The practical change is being able to discuss an ongoing task without repeatedly returning to the chat box. Cursor says users can ask about a diff, check a run, search earlier conversations with that agent or move a call between Project workers.
How Cursor voice chat works
A voice agent handles the conversation and passes coding requests to the agent in the chat. That coding agent keeps its selected model, according to Cursor’s documentation. Choosing a voice therefore does not replace the model doing the programming.
The coding agent delivers brief spoken updates rather than routinely reading its full written responses aloud. Existing dictation remains available for people who want to speak a prompt, inspect the text and send it themselves.
That distinction matters for precise work. Dictation gives the user a chance to catch a misunderstood filename before submitting the request. A live conversation is better suited to exploring a change, asking a follow-up or explaining why the first approach missed the point. These are workflow implications, not measured productivity results.
Starting a call and checking limits
Open an idle agent chat, clear the input and select the waveform button. In the IDE, the chat needs at least one message. A signed-in Cursor account and microphone permission are required. Sessions authenticated with an API key cannot start a call. The button is unavailable in Claude Code chats, archived chats and chats owned by another teammate.
Cursor documents a daily call limit without giving a numerical allowance. By default, calls end after roughly ten minutes of user silence when no work is running. Cloud-agent call transcripts sync across devices during a call unless Privacy Mode is enabled.
For confidential projects, review that setting before the first conversation. Cursor’s data-use policy says Privacy Mode prevents training on customer data and provides zero-retention agreements with model providers. It also describes exceptions for abuse investigations. Turning the setting off allows Cursor to store and use code-related data to improve its features and train models.
Voice access does not settle the coding bill
Neither the launch notice nor the voice guide specifies a per-minute price or a complete plan-by-plan call allowance. Cursor’s current pricing page describes shared AI usage across Cursor and Grok, with weekly refreshes and optional top-offs. It separately says third-party models in Cursor are billed by usage and excluded from the displayed monthly subscription prices.
Check both the available voice allowance and the coding model’s charges before planning long sessions. The ability to keep talking while an agent works should not be read as a promise of unlimited calls or included programming usage.
Keep the result in view
Cursor has also been expanding the ways agents present their work, including cloud-agent canvases and programmable environments. Voice adds a conversational way to supervise that work.
The sensible first test is a small change with a clear expected result. Ask the agent to explain its approach aloud, then inspect the actual diff and test output before accepting the work. A concise spoken update can help a developer stay oriented. The files still show what changed.
