--- title: "Voice Coding on Mac Without the Cloud: Dictate into Cursor, VS Code, and Terminal" description: "Push-to-talk dictation for developers on Apple's Neural Engine. 140ms, private, $10/month Pro. Plus over 300 MCP tools for system control from Claude Code." date: 2026-04-05 author: "Ben Racicot" tags: ["Voice Coding", "Developer Tools", "Speech to Text", "Privacy", "macOS", "On-Device AI", "Cursor", "VS Code", "MCP", "Vibe Coding"] type: "article" canonical: "https://modelpiper.com/blog/voice-coding-mac-local/" --- # Voice Coding on Mac Without the Cloud: Dictate into Cursor, VS Code, and Terminal > Push-to-talk dictation for developers on Apple's Neural Engine. 140ms, private, $10/month Pro. Plus over 300 MCP tools for system control from Claude Code. ## TL;DR "Vibe coding" took off in 2025, but every popular dictation tool sends your audio and screenshots to cloud servers. If you're dictating code, that means proprietary source, internal APIs, and env variables are all passing through a third party. ToolPiper Pro runs push-to-talk on the Neural Engine at 140ms, $10/month. Plus over 300 MCP tools let you control your Mac from Claude Code or Cursor. "Vibe coding" entered the mainstream in 2025. Speak to your AI coding assistant instead of typing. Describe what you want and the AI writes the code. Wispr Flow has built an entire marketing vertical around it - dedicated landing pages for Cursor, Windsurf, and Replit, blog posts about voice-driven development, case studies claiming significant productivity gains. The pitch is compelling. The implementation has a problem. ## What's on your screen when you code When you use Wispr Flow to dictate into your IDE, every word you speak is sent to OpenAI or Meta servers. Wispr also captures screenshots of your active window for "context awareness" - formatting dictation based on what app you're using. Think about what's visible on your screen during a typical coding session. Proprietary source code. Environment variables with API keys. Internal API endpoints and auth tokens. Database schemas. Architecture docs. Slack threads about unreleased features. Pull request reviews with security-sensitive changes. Terminal output with server configs. All of that is in the screenshot that goes to cloud servers alongside your audio. You're not just dictating text. You're sharing your entire working context with a third party's infrastructure. For open-source work, maybe that's fine. For proprietary codebases, startups with unreleased products, or any company with an NDA, it's a data exposure most security teams wouldn't approve if they knew it was happening. ## What developers actually need from voice input Strip away the marketing and there are two core use cases. First, dictation. Speaking code comments, docs, commit messages, PR descriptions, Slack messages, and natural language prompts for AI assistants. This is the 80% case. You're not speaking Python syntax - you're speaking English that happens to be about code. Second, system control. "Run the tests." "Open the browser." "Switch to dark mode." "Mute my Mac before this meeting." These are system commands with nothing to do with text editing and everything to do with developer workflow. Wispr Flow handles the first case via cloud processing. It doesn't handle the second at all. ToolPiper handles both, locally. ## Push-to-talk in your IDE [ToolPiper](https://modelpiper.com) provides push-to-talk dictation that works in every text field on your Mac, including every IDE. Hold Right Option, speak, release. Text appears at your cursor. The Parakeet STT model runs on the Neural Engine at roughly 140ms end-to-end. No audio leaves your Mac. Where this changes a developer's workflow. Code comments. Hold the key, describe the function, release. A well-written comment appears inline. You're more likely to actually write comments when speaking them takes less effort than typing them. Commit messages. Hold the key while looking at the diff, describe what changed, release. Speaking a commit message produces better descriptions than the "fix stuff" you type when you're in a hurry. PR descriptions. Hold the key, explain the PR, release. Spoken explanations tend to be more thorough because speaking is lower friction than typing. AI prompts. When using Cursor or Claude Code, hold the key and describe what you want. Speaking a multi-sentence prompt is significantly faster than typing one. The AI gets a clearer instruction because you explained it naturally instead of abbreviating to save keystrokes. ## MCP tools: system control for your AI workflow This is what no cloud dictation tool offers. ToolPiper exposes over 300 MCP tools covering 142 macOS system actions across 26 domains. If you use Claude Code, Cursor, or Windsurf, these integrate directly into your AI workflow. ``` claude mcp add toolpiper -- ~/.toolpiper/mcp ``` After that one-line setup, your AI assistant can mute your Mac, switch to dark mode, snap windows to specific positions, open files in Finder, adjust display brightness, toggle Do Not Disturb, list running apps - all through natural language prompts alongside your code work. "Mute my Mac, go dark, set brightness to 30%" is three system actions dispatched in sequence from a single prompt in your development environment. Wispr Flow's Command Mode can rephrase a paragraph. ToolPiper's MCP tools can rearrange your workspace while you keep coding. ## Push-to-command: voice-driven system control Even without an MCP client, ToolPiper's push-to-command mode gives you voice control. Hold Right Command, speak an instruction, release. A local LLM interprets your command against 26 action domains and your Mac executes it. A notification confirms what happened. "Open Terminal." "Snap VS Code to the left, Safari to the right." "Mute, go dark, set brightness to 30%." "Play." "Pause." "Lock my screen." Every command runs locally. The STT runs on the Neural Engine. The LLM runs on the Metal GPU. No internet required. ## Accuracy for code dictation Developers have legitimate concerns about STT accuracy for technical content. Code-related dictation includes variable names, framework names, and mixed English-code phrasing ("add a useEffect hook that calls fetchUsers on mount"). Honest assessment - cloud models with billions of parameters do handle niche technical vocabulary better than local models. If you frequently dictate highly specialized terms, a cloud model will more often get the exact phrasing right on the first try. But for the actual use cases developers care about - comments, commit messages, PR descriptions, Slack messages, AI prompts - Parakeet's accuracy is more than sufficient. You're speaking English sentences about code, not dictating regex patterns. The few corrections you occasionally need take less time than the network round-trip you avoid by processing locally. Push-to-talk is additive. It's another input mode, not a replacement for your keyboard. Use it when speaking is faster (descriptions, explanations, messages) and type when precision matters (variable names, command syntax). ## Comparison ToolPiper Wispr Flow Apple Dictation Price Free $12/month Free Processing On-device (Neural Engine) Cloud (OpenAI/Meta) Cloud or on-device Screenshots sent to cloud No Yes No Push-to-talk in IDE Yes (Right Option) Yes (configurable) fn fn MCP tools for AI assistants 29 tools, 142 actions None None System commands by voice 26 domains Text editing only None Offline Yes No On-device mode only ## Setup for developers Install [ToolPiper](https://modelpiper.com/toolpiper) from modelpiper.com. Download [ToolPiper](https://modelpiper.com) from modelpiper.com. Grant Accessibility permission. For MCP integration with Claude Code: ``` claude mcp add toolpiper -- ~/.toolpiper/mcp ``` Hold Right Option in your IDE and start speaking. Hold Right Command to control your Mac by voice. Your code, your audio, and your working context stay on your machine. _ToolPiper is part of the [ModelPiper](https://modelpiper.com) family of local AI tools for Mac. See also: [Wispr Flow Alternative](/blog/wispr-flow-alternative-free-mac), [Push-to-Talk AI on Mac](/blog/push-to-talk-ai-mac), [Desktop Automation with AI](/blog/desktop-automation-ai-mac)._ ## FAQ ### Does ToolPiper understand programming terminology? Parakeet handles common programming terms well - function, component, API, endpoint, repository, deployment, etc. For highly specialized or unusual variable names, you'll occasionally need to correct. But you're typically dictating comments and descriptions in English, not raw code syntax. ### Can I use ToolPiper and GitHub Copilot / Cursor at the same time? Yes. ToolPiper provides voice input (push-to-talk) and system control (MCP tools). Copilot and Cursor provide code completion and AI chat. They serve different functions and don't conflict. ToolPiper's MCP tools actually integrate with Cursor's MCP support, so your AI assistant can dispatch system actions. ### How does the MCP integration work with Claude Code? One command: claude mcp add toolpiper -- ~/.toolpiper/mcp. After that, all 29 action tools appear in Claude Code's tool set. You can ask Claude to mute your Mac, manage windows, toggle system settings, and more - all from the same prompt where you're discussing code. ### Is there a VS Code extension? ToolPiper doesn't need one. It registers global hotkeys at the system level, so push-to-talk works in VS Code (and every other app) without any extension. Hold Right Option in VS Code and speak - text appears at your cursor. The MCP integration works through ToolPiper's MCP server, not a VS Code extension.