# Cursor Voice Chat: How to Start a Call and Fix a Missing Button Cursor now supports spoken conversations with coding agents, but the waveform button, dictation microphone and account requirements do different jobs. By Paras Madan | 2026-11-07 Cursor's new Voice Chat lets you talk with an agent and hear its replies while coding work continues. To start, open an eligible chat, leave the input empty and wait until the agent is idle, then use the waveform button. The nearby microphone starts dictation instead. Cursor [released Voice Chat on November 6](https://cursor.com/changelog/voice-chat) for its Agents Window, IDE, web interface and iOS app. Grok Voice handles the conversation, while the coding agent keeps the model you selected. Talking to the agent therefore does not require moving your coding work to a different model. The useful question is where speaking saves effort without making the task less precise. Explaining why a feature feels wrong can be easier aloud. Checking whether an authorization condition changed still deserves a careful look at the code. ## Voice Chat versus dictation With dictation, your speech becomes text that you can inspect before sending. Voice Chat is an ongoing conversation that can hand coding instructions to the agent as you talk. Earlier Cursor releases used the name Voice Mode for dictation, which makes older tutorials easy to misread. The [prompting documentation](https://cursor.com/docs/agent/prompting) still describes the microphone as voice input. In the Agents Window, you can hold Ctrl+M to dictate; the IDE has its own voice-input shortcut. Choose this route when you want to speak a long prompt but check names, paths and constraints before the agent receives it. **Use dictation when reviewing the instruction matters more than keeping the conversation moving.** A mistyped component name is visible in the input. An imprecise spoken request can be harder to notice once you have moved into discussing the next change. ## Why the Start Voice Chat button is missing Cursor's [Voice Chat guide](https://cursor.com/docs/agent/voice-chat) gives a specific set of checks: - Leave the input empty and wait for the agent to become idle. - In the IDE, use a chat that already contains a message. - Open your own active chat. Archived chats, Claude Code chats, subagents and Project workers do not show the button. - In the IDE, check that dictation is enabled and that a cloud agent is not still streaming. - Update Cursor if those conditions are satisfied. A visible button does not prove the account is ready. Starting a call requires a signed-in Cursor account; API-key authentication alone does not qualify. Microphone access must also be allowed in the relevant device or browser settings. These are different failure points, so deal with them separately. A missing control is a chat-state question first. A sign-in error is an account question. A blocked microphone is a permission question. Repeatedly changing the coding model is unlikely to clarify any of them. ## A useful workflow for reviewing an agent's changes Consider a hypothetical export feature that should include only the rows a user selected. You could begin by asking the voice agent to explain how the current implementation chooses those rows, without requesting changes yet. That keeps the first exchange focused on understanding the behavior. Then describe the discrepancy in concrete terms: selecting two rows should produce two records, but exporting the entire filtered table would be wrong. Ask the agent to state its interpretation before proposing a fix. This is a suggested workflow, not a report of a hands-on test. Once the requirement is clear, put the acceptance conditions in writing. For this example, they might cover an empty selection, a selection spanning multiple pages and a selection that changes after filtering. Cursor supports attaching relevant files or diffs through its text input, which is useful when a spoken description leaves room for ambiguity. After implementation, compare the actual diff and test output with those conditions. An articulate spoken explanation can make a change easier to follow, but it cannot establish that the change is correct. **Keep the evidence on screen even when the discussion happens aloud.** ## Privacy and call limits need a separate check Cursor says Voice Chat uses zero data retention when Privacy Mode is enabled. Its [data-governance documentation](https://cursor.com/docs/enterprise/privacy-and-data-governance) also distinguishes model requests from Cloud Agents, which need repository access while working. A voice-specific retention statement should not be stretched into a claim that every part of a development workflow runs locally or stores nothing. Check your team's approved configuration before discussing customer details or confidential code. Speaking can feel less formal than pasting a file into a chat, but the sensitivity of the information has not changed. There is a daily calling limit, though the Voice Chat guide does not publish a fixed allowance. It also describes an idle timeout after roughly ten minutes of silence when no work is running. Do not build a workflow around an assumed unlimited call duration. Voice is most useful when it makes it easier to explain intent and question an implementation. Keep exact requirements, review decisions and test evidence readable afterward, so the convenience of the conversation does not become ambiguity in the code. I'm Paras. I work at the intersection of AI, engineering and distribution. I help technical founders explain developer tools through clear guides, useful examples and product stories people can actually use. [Book a 30-minute call](https://cal.com/paras-madan/30min).