Developer guide
Voice coding in 2026: dictate prompts to Claude Code, Cursor, and every other app — offline
If you spend your day talking to coding agents — prompts, PR descriptions, commit messages, Slack updates — you now write more prose than code. This page covers the built-in voice modes in Claude Code and Cursor, exactly where they stop, and how to dictate into everything without your audio leaving your machine.
The shift
AI coding turned developers into prose writers
The work changed. A detailed 150-word prompt gets a coding agent much further than "fix the bug" — and the same goes for the PR description, the commit message, and the Slack thread explaining what shipped. Most developers type 40–60 words per minute. Speech runs roughly 150–220. For the sentence-shaped work that AI workflows multiplied, dictation is a 3–5x speedup.
And that work doesn't live in one app. The same paragraph-writing happens in Claude Code, in Cursor's chat panel, in a GitHub PR form, in a ticket, in Slack. Every voice feature shipping inside an AI tool covers exactly one of those surfaces. That's the gap a system-wide dictation tool fills: hold a hotkey, talk, release — the text lands wherever your cursor is.
The built-ins
Claude Code's voice mode is good. It's also a silo.
Claude Code added voice dictation in early 2026, and inside its own walls it's genuinely good. Run /voice, hold Space to talk (there's a tap-to-record mode too), and it understands coding vocabulary — it even feeds your project name and git branch to the recognizer as hints, and transcription doesn't consume tokens. If you live in the Claude Code CLI or its VS Code extension, it's worth turning on.
The limits are documented by Anthropic itself. The docs state that voice mode streams your recorded audio to Anthropic's servers and that "Audio is not processed locally." It requires a Claude.ai sign-in, and is unavailable if you authenticate with an API key, Bedrock, or Vertex AI, if your org has HIPAA compliance enabled, or in SSH sessions, dev containers, and Claude Code on the web. And it works only inside Claude Code — switch to Cursor, your browser, or Slack, and you're back to typing.
Cursor is the same story, narrower. Voice input arrived in Cursor 2.0 (October 2025) for the chat input only, and the June 2026 release added voice narration to Design Mode — still nothing for the terminal, PR descriptions, or issues. VS Code's free Microsoft Speech extension is an honorable exception (it processes audio locally), but it works only inside VS Code. Three tools, three partial voice features, none of them system-wide.
Why cloud audio matters more for developers
A dictated prompt isn't just your voice. It's you reading out function names, schema details, unreleased feature names, customer context — whatever the agent needs to know. With cloud dictation, all of that rides along with your audio to a third party's servers. With SuperDictate, the speech model runs on your device: audio is processed 100% on-device and never uploaded. Works in airplane mode, and on NDA codebases without a meeting with your compliance team.
Side by side
SuperDictate vs the built-in voice modes
| SuperDictate | Claude Code voice | Cursor voice | |
|---|---|---|---|
| Where it works | Every app on Mac and Windows — Claude Code, Cursor, terminal, browser, Slack, GitHub PRs | Claude Code CLI + its VS Code extension only; not over SSH, dev containers, or Claude Code on the web | Chat input only, plus Design Mode narration (June 2026); not the terminal, PRs, or issues |
| Audio processing | 100% on-device, never uploaded — works in airplane mode | Streamed to Anthropic's servers; docs: "Audio is not processed locally" | Cloud transcription |
| Requirements | macOS 13+ or Windows 10/11 | Claude.ai sign-in; blocked for API-key, Bedrock, Vertex, and HIPAA orgs | Cursor subscription |
| Works offline | Yes | No | No |
| Price | $79 once — Mac + Windows; free 15 min/month | No extra charge on eligible Claude plans — stop subscribing, lose it | Included with Cursor's subscription |
Claude Code and Cursor details verified June 2026 from their public docs and changelogs; check their pages for changes.
Credit where due
Talon is for code. SuperDictate is for prose. Run both.
If you want to write and edit code itself by voice — speak syntax, navigate symbols, drive the whole OS hands-free — the answer is Talon, not SuperDictate. Talon's public release is free on Mac, Linux, and Windows, it processes speech locally, and it's the standard for command-and-control voice coding, especially for developers managing RSI. The trade-off is real: a phonetic alphabet and a command grammar that take weeks to internalize.
SuperDictate is the other half of the workflow: natural prose. Speak a paragraph, get clean text — no grammar to learn. Plenty of Talon users pair it with a prose dictation tool for exactly this reason: commands for the code, dictation for the words around the code. (Serenade, the old natural-language coding tool, is effectively dormant — original development stalled years ago, and a community fork shipped its first beta in January 2025.)
One more honest note on the local-dictation field: most of it is Mac-only. EmberType ($49 one-time) requires macOS 14+ on Apple Silicon, Spokenly is Mac and iOS, and VoiceInk is Mac-only. superwhisper did add Windows support in December 2025 — but its lifetime tier is $249.99, against $79 here. If half your life is a Windows machine, your options shrink fast — that's the case SuperDictate was built for, and why one license covers both. See the full field on our Windows dictation roundup. (Prices and platform support as of June 2026.)
Setup
Dictate your first prompt in five minutes
Install SuperDictate
Download for Mac (macOS 13+) or Windows (10/11). Free to try — 15 minutes every month. No credit card.
Grant permissions, pick a hotkey
On Mac, allow the microphone and Accessibility (that's how it types into other apps); on Windows, just the microphone. Set your push-to-talk key.
Click into Cursor's chat and talk
Put your cursor in any text field — Cursor chat, the Claude Code terminal, a PR description, Slack. Hold the hotkey, speak your prompt, release. The text appears where your cursor is.
Optional: turn on AI polish
Off by default. When enabled, it cleans up the transcript — text-only, via local Ollama or your own API key. Your audio still never leaves your computer.
Questions
Voice coding, answered
Doesn't Claude Code already have voice dictation built in?
Yes. You can enable it with /voice and hold Space to talk, and it is genuinely good with coding vocabulary. But it only works inside Claude Code (the CLI and its VS Code extension), it requires signing in with a Claude.ai account — it is unavailable on API-key, Bedrock, or Vertex AI setups — and Anthropic's docs state that audio is streamed to their servers, not processed locally. SuperDictate types wherever your cursor is — Claude Code, Cursor, your browser, Slack — and your audio never leaves your computer.
How do I dictate into Cursor?
Cursor's built-in voice input (added in Cursor 2.0) only works in the chat input — not the terminal, PR descriptions, or issues. A system-wide dictation tool covers all of them: hold a hotkey, speak, and the text lands wherever your cursor is. SuperDictate does this on both Mac and Windows, and the speech model runs on your device.
Is SuperDictate a replacement for Talon?
No — they solve different problems. Talon is command-and-control voice coding: you learn a spoken command grammar to write syntax, navigate files, and drive your computer hands-free. It is excellent, and the public release is free. SuperDictate is prose dictation: speak naturally and get clean text for prompts, PR descriptions, commit messages, and Slack replies. Many Talon users run a prose dictation tool alongside it.
Does my voice or my code context go to the cloud?
No. Audio is processed 100% on-device and never uploaded — SuperDictate works in airplane mode. The optional AI polish step is text-only, off by default, and runs via local Ollama or your own API key. That is the difference from Claude Code's built-in voice mode, which streams audio to Anthropic's servers, and from cloud dictation tools.
Why dictate prompts to an AI agent instead of typing them?
Speech runs roughly 150–220 words per minute; most developers type 40–60. Detailed prompts give coding agents more context to work with, and speaking turns a 150-word prompt from a three-minute typing job into about a minute of talking. Free to try — 15 minutes every month. No credit card.
Does one license really cover both my Mac and my Windows PC?
Yes. Pay $79 once, own it forever. One license covers your Mac and your Windows PC. Most local dictation tools in this niche — EmberType, Spokenly, VoiceInk — are Mac-only, so Windows developers are usually pushed toward cloud subscriptions. SuperDictate runs on macOS 13+ and Windows 10/11.
Talk to your agent. Keep your audio.
Dictate prompts into Claude Code, Cursor, and every other app. Free to try — 15 minutes every month. Pay $79 once, own it forever, on Mac and Windows — with a 30-day money-back guarantee.
Free to try — 15 minutes every month · No credit card · 30-day money-back guarantee