In agentic coding, you spend more time telling an AI what to build than writing code yourself. Claude Code, Codex, Cursor and Windsurf do the implementation; you describe, redirect and review. Most of your input is now plain language, and plain language is what voice is good at.
Typing speed used to be a coding skill. With agents, what matters is how quickly and clearly you can explain a change. Conversational speech is commonly around 130 to 150 words per minute, well above most people's typing.
Why voice works here#
Speed
A feature description that takes a couple of minutes to type takes well under a minute to say. Agents get dozens of these a day.
Better prompts
Typed prompts get terse. Spoken ones include the context, constraints and edge cases an agent needs.
Less fatigue
Long prompting sessions are a lot of typing. Speaking moves most of it off your hands.
Review is conversation
"Good, but make the error handling more specific" is easier to say than to type.
The voice and agent loop#
Each step is the same action in Whisperer: hold Fn, speak, release, press ⌘V, send. Whisperer copies the text and shows a Copied ⌘V toast; it never types or presses Return for you.
1. Initial direction#
“create a REST API for a todo list using Express and TypeScript with CRUD endpoints input validation with Zod error handling middleware and Jest tests for each endpoint”
Create a REST API for a todo list using Express and TypeScript, with CRUD endpoints, input validation with Zod, error handling middleware, and Jest tests for each endpoint.2. Course correction#
“the error middleware should return JSON with an error code a message and an optional details field don't use the default Express handler”
The error middleware should return JSON with an error code, a message and an optional details field. Don't use the default Express handler.3. Review by voice#
“this looks right but use a connection pool instead of opening a connection per request and add a graceful shutdown handler”
This looks right, but use a connection pool instead of opening a connection per request, and add a graceful shutdown handler.4. Approve and commit#
“this is good commit it with the message add todo API with CRUD endpoints and validation”
This is good. Commit it with the message: Add todo API with CRUD endpoints and validation.Setting up#
Install Whisperer
From the Mac App Store, $2.99. Apple Silicon and macOS 26 Tahoe.
Add your vocabulary
Library, service and product names go in the personal dictionary, so "Zod" and "PostgreSQL" don't come out as something else.
Use a clear mic
Built-in is fine in a quiet room. A headset helps when you're giving long, detailed instructions.
Watch the live preview
Words appear as you speak, so you can catch a misheard name before you paste.
There's nothing to set per tool. Whisperer writes what you say as plain text, which is exactly what agents take. It doesn't do spoken symbol or casing commands; the agent writes the syntax.
Give agents what you've said, not just what you type#
This is the part that changed in Whisperer 2.0. With Pro, Whisperer can run a local MCP server, so your agent can read your meetings and dictations directly:
search_meetingsandget_meeting_transcriptlet it find the call where a feature was discussedsearch_transcriptionsandget_transcriptionlet it read a long dictation you made earlier, such as a spoken bug report or design notesadd_meeting_notelets it write a summary back
One-click setup covers Claude Code, Codex, Cursor, VS Code and Windsurf, among others. The server is off by default and listens only on your Mac. A prompt like "find what we decided about rate limits in Monday's call and implement it" then works from the real conversation. See the MCP page and meeting notes.
Tools that work well with voice#
| Tool | Voice use |
|---|---|
| Claude Code | Spoken instructions in the terminal; reads your meetings via MCP |
| Codex | Spoken tasks in CLI, IDE or app; MCP supported |
| Cursor | Agent and chat prompts, inline edit prompts, review |
| Windsurf | Agent prompts and multi-file changes |
| VS Code with Copilot | Chat and agent prompts, commit messages |
Where this is going#
More coding happens through conversation with agents and less through typing syntax. You don't need to dictate everything. But if you spend hours a day prompting AI tools, speaking those prompts is worth a week's trial.
Related: Voice dictation for vibe coding, Voice coding guide, Claude MCP for meeting notes, Developer productivity with voice. See pricing.
Meeting notes and dictation that run on your Mac
$2.99 app, Pro $14.99, paid once. No account, no subscription.
Get Whisperer