AI Writing with Voice

Voice Input for Codex, Claude Code, and Cursor

For voice input in Codex, use the microphone in the desktop app's prompt box. In the Claude Code CLI, run /voice hold and hold Space while you speak. In Cursor, click the microphone in the Agent chat input. Each route gives you text to review before sending when you use the controls below. OpenAI's Codex dictation demo · Claude Code voice dictation · Cursor Agent voice input

Codex voice input in the desktop app

Open a Codex chat in the ChatGPT desktop app and use the microphone in the composer to dictate the task. Stop recording, read the text, and send when the scope is right. The app also has a Voice control for a live conversation that can start or coordinate work. Use dictation when you need to inspect a prompt before it becomes a task. OpenAI's Codex dictation demo · OpenAI's guide to Voice with Codex

Claude Code voice input in the terminal

In a local Claude Code CLI session signed in with a Claude account, run /voice hold. Hold Space until the footer shows that it's listening, speak, then release to finish. With the default hold-mode settings, the prompt stays in the input until you press Enter. The CLI also has a tap mode that can send as soon as you stop recording. Claude Code streams audio to Anthropic for transcription. If native dictation is unavailable in an SSH session or with API-key authentication, dictate into a local editor and paste the text at Claude Code's prompt. Claude Code requirements and recording modes

Cursor voice input in Agent chat

Click the chat input's microphone in the IDE or Agents Window. In the Agents Window you can also hold Ctrl+M while speaking. When the transcript appears, use the keyboard to fix paths or attach the files you want the agent to inspect. The waveform control starts a two-way voice chat, where the voice agent can hand work to the coding agent during the conversation. Cursor Agent voice input · Cursor voice chat

Turn your voice into a coding prompt the agent can verify

Say the problem and the outcome in your own words. Then paste literals such as file paths, flags, and error strings from the editor or terminal. An example of the prompt after review follows.

Investigate why checkout retries create a second order when the first request times out. Start in src/checkout/submitOrder.ts and look for DUPLICATE_ORDER in the logs. Explain the cause and suggest a regression test that simulates a timeout followed by a retry. Wait for my go-ahead before editing files.

The part worth speaking is the sequence of events. The exact file name and error string are easier to paste. Before you send, check both and make sure the request says whether the agent should investigate, propose a change, or edit files.

Setup guide

Voconly · Use one dictation shortcut across coding tools

On Windows x64 or a Mac with Apple silicon, set up a local speech model. Draft into a focused chat field or a plain text editor, then check the paths and paste the finished prompt into your agent.

Save 30 minutes of typing every day

Local processing, offline ready, data safe

More Downloads

Windows · macOS dual platform