Voice coding with AI means speaking your instructions to a coding agent and letting the agent write the code, instead of typing every prompt. It works because good prompts are long: the context, the constraint, the edge case you just remembered. Saying that out loud takes seconds. In AgentVera every agent’s composer has a mic button that transcribes your speech locally with Whisper, puts the text in the box for you to fix, and sends it when you are ready.
This post covers why voice fits agent work, how dictation in AgentVera works, and habits that make spoken prompts land.
Why voice fits coding agents
Classic voice coding meant dictating syntax: “open paren, const, x, equals.” That never caught on for most people. Coding agents change the job. You no longer type the code; you describe the change, and describing is what speech is good at.
| Typing a prompt | Speaking a prompt | |
|---|---|---|
| Long context | Tempting to cut short | Easy to include |
| Speed | Your typing speed | Usually faster than typing |
| Hands | On the keyboard | Free to scroll, point, sketch |
| Precision | Exact names and paths | Names may need a quick fix |
| Best for | Short, exact commands | Explaining intent, reviewing out loud |
The weak spot is precision. Speech recognition can mishear a function name or a file path. That is why dictation in AgentVera puts the text in the box first instead of sending it straight away.
How dictation works in AgentVera
AgentVera is a desktop app for macOS, Windows and Linux that runs coding agent CLIs side by side. Each agent has a composer, the input box under its terminal, and the composer has a mic button with the tooltip Dictate (local Whisper).
Steps
- Click the mic button in an agent’s composer.
- Speak your prompt.
- Click again (Stop recording and transcribe). The button shows Transcribing… while it works.
- The text appears at your cursor in the composer.
- Fix any misheard name or path, then send.
If you said nothing, you get “No speech detected.” and nothing is inserted. Because the text lands where your cursor is, you can type the exact part (say, a file path) and speak the rest around it.
Local Whisper, nothing sent to a server
Transcription runs on your own computer with whisper.cpp. Your recording is not uploaded anywhere, and the temporary audio file is deleted after it is transcribed. Only the text you send reaches the agent.
| Question | Answer |
|---|---|
| Engine | whisper.cpp (whisper-cli) on your machine |
| Setup | The packaged app ships with whisper-cli and the small model |
| No model found | AgentVera offers to download the base model, about 142 MB, once |
| Language | The app’s interface language, English or Turkish |
| Hardware | Uses Metal on macOS; on x64 machines the bundled build needs a CPU with AVX2 |
| Price | Part of the free plan |
If you already have whisper.cpp installed and it is on your PATH, AgentVera can use it. When both a small and a base model are present, it prefers the small one, which is more accurate.
Works with every agent
The mic button sits in the composer, not inside a specific CLI, so the same dictation works for Claude Code, Codex, Gemini CLI and any other agent you run. The composer’s other tools still apply: arrow keys bring back earlier prompts, drafts are kept, and you can queue a prompt while the agent is busy.
First-time setup and troubleshooting
The first time you click the mic, your operating system asks whether AgentVera may use the microphone. Allow it. If no speech model is found, AgentVera asks Download the speech model? and explains that the whisper.cpp base model (about 142 MB) goes into AgentVera’s data folder. Choose Download model and the composer shows the progress as a percentage. After that, dictation works offline.
| What you see | What to do |
|---|---|
| Could not open the microphone | Grant microphone access to AgentVera in your system settings, then try again |
| No speech detected. | Check the input device and speak a little closer, then record again |
| whisper.cpp not found | Outside the packaged app, install whisper.cpp: brew install whisper-cpp on macOS, a release package on Windows, your package manager on Linux, with whisper-cli on your PATH |
| The recording is invalid or too long | Split a very long prompt into two recordings |
| Wrong language in the text | Switch the app’s interface language to the one you speak |
Voice coding from your phone
The AgentVera mobile app for iOS and Android connects to your computer, and its Chat view has dictation too: you can speak for up to 60 seconds. The audio is transcribed on your computer, in memory, without being written to disk. That is handy when an agent asks a question while you are away from the desk; see the Claude Code mobile app for the full remote setup. If you steer agents from Telegram instead, note that voice messages are not supported there; type your reply. More in controlling Claude Code from Telegram.
How to write prompts by voice
Speaking a prompt is different from typing one. These habits help:
- Say the goal first. “I want the signup form to show a password strength hint.” Then the details.
- Name the place. “In the signup page component, under the password field.” If the exact file name matters, type it into the text after dictating.
- Add the constraint you would forget to type. “Don’t add a new dependency, and keep the existing validation.”
- Say how to check it. “Run the tests for the form when you’re done.”
- Read before sending. One glance catches the misheard word.
A spoken prompt often ends up longer and more complete than a typed one, and that is the point: agents do better with context. How to vibe code has more on prompts that keep an agent on track.
Good moments for voice
- Reviewing a diff. Scroll through the changes and say what to fix as you go.
- Explaining a bug. Describe what you saw, what you expected and what you already tried.
- Planning. Talk through a feature and ask the agent to turn it into steps before it writes code.
- Long sessions. Give your hands a break from typing.
When to type instead
- Exact identifiers, commands and paths.
- Short commands like
/compactor a one-word answer. - Shared spaces where talking out loud is awkward.
Quick checklist
- Microphone permission is granted to AgentVera.
- The interface language matches the language you speak.
- The speech model is downloaded, or the bundled one is in place.
- You read the text before sending it.
Download AgentVera to talk to your agents, or read more on the voice dictation page.