Voice coding with AI agents: say what you want, not how to type it
Voice coding with AI agents: describe intent instead of dictating syntax. Hear a 30-second hands-free session with Claude Code and the tools behind it.
For years, voice coding meant dictating code itself: every bracket, every camelCase name, every cursor move had a spoken command. It worked, and for developers who can't type all day it was a lifeline, but it took weeks to learn and it was slower than a keyboard for most people.
AI coding agents change what you say. You no longer speak the syntax; you speak the intent: "add a health endpoint that returns the uptime, then run the tests". That's a sentence, and sentences are what voice is good at.
Voice coding before agents
The long-standing tool is Talon, which maps a spoken vocabulary, often with noises and eye tracking, onto keystrokes, editor commands and code. It's powerful and deeply customisable. It's also a new language to learn, because you're still steering the editor one action at a time.
That model fits code you write yourself. With an agent writing the code, most of those commands become unnecessary.
What a voice workflow with an agent needs
Dictation alone covers the first step. A working session with an agent has more moments where your hands would otherwise go back to the keyboard:
- Starting: a wake word, so you don't have to hold a key.
- Speaking the prompt: accurate transcription, ideally in your own language.
- Hearing the result: a short spoken summary when the agent is done, so you don't have to read the screen.
- Deciding: answering permission prompts by voice ("approve", "deny", "explain").
- Steering: checking status, interrupting, and switching between several agents.
The options today
- Claude Code's built-in voice dictation. Type
/voice, then hold Space (or tap) to dictate into the prompt. It works in the terminal and in Anthropic's VS Code extension, understands 20 languages, and needs a claude.ai sign-in. Audio is streamed to Anthropic's servers for transcription. It covers step 2. - Your operating system's dictation. macOS Dictation and Windows voice typing type anywhere, including a prompt box. Also step 2, without any idea of what an agent is.
- Verbative, a VS Code and Cursor extension that wraps the whole loop around the Claude Code CLI: the wake word, on-device transcription in about 99 languages, spoken summaries, permissions by voice, status checks, and switching or creating agents by voice. Speech recognition and the voice both run on your machine.
A hands-free session, step by step
This is what a session with Verbative sounds like:
- "Listen." The agent's prompt goes live.
- "Add a health endpoint to the API that returns OK and the uptime, then run the tests." The transcript types itself in as you speak.
- "Stop listening." The prompt is sent, and Verbative answers "On it."
- Claude asks to run the tests. You hear the request and say "approve".
- While it works, "create agent" opens a second agent; "listen" gives it its own task.
- When each agent is done, you hear what it did: "Done. I added a health endpoint…"
No keyboard, and nothing you'd want to miss on the screen.
Tips that make voice coding work
- Use headphones, so the agent's voice doesn't end up in your microphone.
- Speak like you'd write a ticket: the goal, the file or area, and how you'll know it's done.
- Keep prompts to a sentence or two and let the agent ask for details.
- Cancel freely: with Verbative, "hold on" drops a prompt you don't like before it's sent.
- Speak your own language: Verbative translates the prompt to English on your machine before it reaches Claude.
Frequently asked
Can you code by voice?
Yes, and with AI agents it's far easier than it used to be: instead of dictating syntax, you describe what you want in a sentence and the agent writes the code. A full voice workflow also needs spoken feedback and voice answers to the agent's permission prompts.
What's the best voice coding software for AI agents?
It depends on how much of the loop you want hands-free. Claude Code's built-in /voice is push-to-talk dictation into the prompt. Verbative adds the rest around the Claude Code CLI in VS Code and Cursor: a wake word, spoken summaries, permissions by voice, and switching between several agents.
Is voice coding a good fit for RSI or other typing limits?
Many developers who can't type for long use voice to keep working. With an agent doing the typing, a session can run entirely by voice, from the prompt to the permission prompts to hearing the result.
Does voice coding work offline, on my own machine?
With Verbative, yes: the wake word, the transcription and the spoken voice all run on your machine, so no audio leaves it. Claude Code's built-in dictation sends audio to Anthropic for transcription.