Voice Coding in AI IDEs

Use HyperWhisper with Cursor, Claude Code, and other AI-powered editors to dictate prompts, tag files by voice, and let agents respond to your spoken questions.

HyperWhisper integrates with AI coding agents in two complementary ways: your agent can call HyperWhisper as an MCP tool to transcribe audio files during a session, and HyperWhisper can pop up automatically when the agent pauses to ask you a question or request a permission. On top of that, a dedicated Code preset and Developer Mode help your transcripts land in the right format and carry the right file context before the agent ever sees them.

This page covers the full picture. Deep-dive reference for individual features is linked throughout.

Prerequisites

• HyperWhisper is installed and has been opened at least once.
• Settings → Local API is turned on. This is what writes the discovery file that the MCP bridge reads.
• Node.js 18 or newer is on your PATH (node --version should print a version number). The npx command ships with Node.

The MCP bridge and Claude Code / OpenCode plugins are macOS features. iOS does not currently support editor integrations.

---

Agent integration via MCP

The @hyperwhisper/mcp package is a Model Context Protocol bridge that wraps the local API. Configure it once and any MCP-capable agent — Cursor, Claude Desktop, Claude Code, Zed, custom code — can call HyperWhisper as a tool.

Tools the agent gets

Tool	What it does
health	Confirms HyperWhisper is running.
list_models	Returns the catalog of available voice and post-processing models.
list_modes	Lists your saved transcription Modes.
transcribe	Transcribes an absolute path to an audio file.
post_process	Rewrites text with a preset (hyper, note, email, commit) or a free-form instruction passed via the prompt field.
search_recordings	Substring search across past transcripts.
get_recording	Retrieves a single transcript row by ID.

Setup

Cursor

Edit ~/.cursor/mcp.json and add the hyperwhisper entry:

{
  "mcpServers": {
    "hyperwhisper": {
      "command": "npx",
      "args": ["-y", "@hyperwhisper/mcp"]
    }
  }
}

Restart Cursor. The server should appear under Settings → MCP.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "hyperwhisper": {
      "command": "npx",
      "args": ["-y", "@hyperwhisper/mcp"]
    }
  }
}

Quit and reopen Claude Desktop. The hammer icon in the composer should list hyperwhisper.

Claude Code

Run the CLI helper once:

claude mcp add hyperwhisper -- npx -y @hyperwhisper/mcp

In a new session, /mcp lists hyperwhisper with its tools.

Verify

Open a fresh agent session and ask:

Ask hyperwhisper to run a health check.

The agent should call the health tool and report the app version. If you have an audio file handy, try:

Use hyperwhisper to transcribe /Users/me/Desktop/test.wav

The agent calls transcribe and returns the text inline.

For full tool signatures and troubleshooting, see MCP Setup.

---

Code preset for voice-to-code

When you’re dictating code rather than prose, use a mode with the Code preset. It converts spoken symbol names into code syntax and disables automatic capitalization and punctuation so your identifiers land exactly as spoken.

You say	You get
open paren	(
close paren	)
fat arrow	=>
camel case user name	userName

Set up a Code mode

1. Create a new mode

   Click + in the mode list (or open an existing mode to edit it).

2. Choose the Code preset

   Select Code from the preset picker. This sets the post-processing instructions for symbol conversion.

3. Enable AI post-processing

   Set AI Post-Processing to Cloud or Local — the Code preset requires a post-processing pass to convert spoken symbols.

4. Turn off auto-capitalization and punctuation

   In the mode’s formatting options, make sure Capitalization and Punctuation are off. The Code preset expects raw, case-preserving output.

Consider enabling Screen Text on your Code mode so the AI can see your current file and spell identifiers, class names, and variable names correctly.

---

Developer Mode for file context

Developer Mode lets you tag specific files by voice during a recording. HyperWhisper detects your active IDE, indexes your project files, and resolves spoken file references into proper @filename tags — exactly what AI agents need for precise, context-aware code generation.

Setup

1. Enable Developer Mode

   Open HyperWhisper settings and turn on Enable Developer Mode. See Developer Mode for the exact location.

2. Add your project directories

   Click + Add Directory and select your project root. You can add multiple directories for monorepos or parallel projects.

3. Verify indexing

   The status indicator shows your active project name and a file count. A blue checkmark confirms the IDE is detected.

Using @ mentions

With Developer Mode and AI post-processing active, say "at [filename]" during recording:

You say	Transcript becomes
"at app.swift, add error handling"	@app.swift, add error handling
"at components/button, update the border radius"	@components/button.tsx, update the border radius
"at main.swift and at utils.ts, check the date formatting"	@main.swift and @utils.ts, check the date formatting

HyperWhisper uses fuzzy matching to resolve partial file names and path fragments. The system remembers your last active project for 30 minutes and switches automatically when you change IDEs or projects.

Supported IDEs

Xcode, VS Code, Cursor, Windsurf, IntelliJ IDEA, PyCharm, WebStorm, GoLand, Zed, Atom, and Brackets.

For a deeper look at status indicators and troubleshooting, see Developer Mode.

---

Auto-paste into your editor

Enable auto-paste and HyperWhisper delivers the finished transcript directly into your focused editor window — no manual paste needed.

1. Enable auto-paste

   Open Settings → Text Output and turn on Paste result automatically.

2. Grant Accessibility permission

   macOS prompts you the first time. If you missed it: System Settings → Privacy & Security → Accessibility → enable HyperWhisper.

After a recording finishes, HyperWhisper copies the transcript to the clipboard, focuses your editor, and issues a standard paste. Your previous clipboard contents are restored afterwards.

See Auto Paste & Clipboard Behavior for the full configuration options, including clipboard restore delay.

---

Agent-driven voice responses

Claude Code plugin

The Claude Code plugin registers hooks into Claude Code’s lifecycle. When the agent finishes a task, asks a question, or requests a permission, HyperWhisper opens automatically with the current agent context. Dictate your response — or type it — and HyperWhisper sends it back to Claude Code.

Install:

curl -fsSL https://hyperwhisper.com/install-claude-code.sh | bash

Restart Claude Code after installation.

OpenCode plugin

The OpenCode plugin works the same way for the OpenCode agent: when OpenCode pauses for input, the plugin sends the context to HyperWhisper, you respond by voice, and HyperWhisper writes the response back.

Install:

curl -fsSL https://hyperwhisper.com/install-opencode.sh | bash

Or add to your OpenCode config manually:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["@hyperwhisper/opencode"]
}

---

Workflows

Cursor agent: dictate a prompt

1. Switch to your Code mode.
2. Press your record shortcut, speak your prompt — e.g. “at components/button, refactor the onClick handler to use useCallback”.
3. HyperWhisper transcribes, resolves the @ mention, and pastes the result into Cursor’s composer.
4. Cursor’s agent picks up the prompt with full file context and generates code.

Claude Code: voice-respond to agent questions

1. Install the Claude Code plugin (above).
2. Start a Claude Code session normally.
3. When Claude Code pauses to ask a question or request a permission, HyperWhisper opens.
4. Record your answer. HyperWhisper transcribes and sends the response back to Claude Code automatically.

Standalone transcription via MCP

Any agent session can call transcribe directly on a recorded file:

Use hyperwhisper to transcribe /Users/me/recordings/meeting.m4a and post_process the result with the "note" preset.

The agent calls both tools in sequence and returns a formatted transcript inline.

---

Troubleshooting

MCP server doesn’t appear in the IDE

• Confirm node --version prints 18 or higher.
• Toggle Settings → Local API off and back on, then restart the IDE.
• On first run, npx -y downloads the package — it may take a few seconds on a slow connection.

HYPERWHISPER_NOT_RUNNING error Open HyperWhisper, enable Settings → Local API, then re-run the agent prompt. The bridge reads the discovery file on every request.

Code preset not converting symbols Verify AI Post-Processing is set to Cloud or Local on the mode — the Code preset has no effect when post-processing is off.

@ mentions not appearing in the transcript

• Confirm Developer Mode is enabled (see Developer Mode).
• Click Reindex Files to refresh the file index.
• Ensure AI post-processing is on for the active mode.
• Say "at [filename]" clearly — the fuzzy matcher needs at least a partial filename.

Auto-paste not working

• Grant Accessibility permission in System Settings → Privacy & Security → Accessibility.
• Confirm the IDE window is focused when the recording finishes.
• Try pasting into Notes or TextEdit to rule out an editor-specific issue.

---

Limitations

• The MCP bridge and agent plugins require HyperWhisper to be running — there is no background listener that activates independently.
• Local model transcription followed by local post-processing adds latency; for long code blocks, cloud transcription is faster.
• The Code preset guides the AI’s formatting instructions, but cloud post-processing models may occasionally rephrase output beyond what the preset intends.
• iOS does not support MCP integration, the Claude Code plugin, or the OpenCode plugin.