The shortcut version: don't dictate code itself, dictate the prose around it. Cursor prompts, commit messages, PR descriptions, code comments, the Slack message explaining what you just shipped. That's where Mac dictation pays back hardest for developers.
Verified May 2026 · written by the Murmur team
Every few months someone tries to dictate code directly. The flow goes: "open paren x equals one comma y equals two close paren semicolon." It's exhausting after one line and incomprehensible after three. Speech grammar collapses on nested structures. Punctuation as words is anti-ergonomic. Modal commands ("camel case," "function," "new line indent") help but require a phrase-book mindset that breaks flow.
The conclusion most developers land on: let the keyboard handle code, use dictation for the prose. The wins stack up:
Cursor's chat panel is where most of a modern coding day happens. The prompts are speech-shaped (rough requirements, "make this function async," "add error handling for the timeout case"), and they benefit from polish before sending so Claude has a clean spec.
The Murmur-Cursor flow:
The friction this removes is small per turn (5-10 seconds) but you do this hundreds of times a day. The compound saving is real, and the prompt quality is higher because Claude saw the file instead of needing it described.
Claude Code is text input in a terminal. Mac dictation pastes wherever the cursor is, including Terminal, iTerm, Ghostty, and Warp. The flow is identical to Cursor's chat panel, just at the command line.
One small workflow tweak for terminal-heavy use: bind Murmur's hotkey to something reachable from the typing position (left Control, right Option, or a function key). The mouse-side default works for GUI apps but is slightly more awkward when your hands haven't left the keyboard.
Silent mode is particularly good for one-shot terminal commands. Long-hold the hotkey, mumble "find all .tsx files modified in the last week," let go. Murmur runs Claude as a quick AI assistant and pastes the actual command. Hit enter.
Screen-context AI is the killer feature for coding. Claude sees the file when polishing, so prompts include context without being typed. Audio stays local, which matters for proprietary code under NDA.
Run the LLM step on a local Ollama model and nothing leaves your Mac. Right pick if your security profile rules out cloud LLM calls even for transcribed text.
Polish quality is excellent across Slack, Linear, GitHub, your IDE. Audio is uploaded to their cloud, which can be a non-starter for code under NDA. Read your contract before using it for work.
The dictation patterns that pay back fastest in a coding day:
Bug report → reproducible STR. "Steps to repro: open the editor, paste a JSON blob bigger than 5MB, click save, watch it return a 500. Expected: should chunk or stream. Actual: dies with 'maximum payload exceeded' at line 47 of writeBuffer." Long enough to be tedious to type, structured enough to dictate cleanly.
Cursor prompt → spec. "Add error handling for the timeout case in the request hook. Use AbortController if not already, and show a toast on failure with a retry option. Don't change the success path." Hand-typing this is 30 seconds; dictating + AI polish is 5.
PR description → bullets. "Three things in this PR: first, the timeout bug fix from issue 412. Second, a small refactor to extract the toast component because we'll reuse it next sprint. Third, a test for the abort case." Murmur's AI mode formats this into Markdown automatically.
Commit message → why-not-what. "Refactor request hook to share AbortController between the auth and data fetches so we don't get orphaned requests on logout." Speak the why; the AI shapes the format.
You can, but you shouldn't. Speaking syntax (open paren, x equals one, close paren) is slower than typing. The high-leverage use of dictation is the prose around code: prompts to Cursor or Claude Code, git commit messages, PR descriptions, comments, Slack messages about what you're working on.
Murmur ($29 lifetime). The screen-context AI mode takes a screenshot of Cursor before polishing, so Claude sees the file you're on, your selection, and the surrounding code. You mumble what you want changed; Murmur pastes a polished prompt that already includes context.
Yes. Claude Code is just text input in your terminal. Dictation pastes wherever the cursor is. Murmur, Wispr Flow, and Apple Dictation all work into Terminal, iTerm, Ghostty, and Warp.
With Murmur Quick mode, no. Audio and transcription stay local. With Murmur AI mode, transcribed text plus a screenshot go to Claude via your own API key. With Wispr Flow / Aqua Voice, audio is uploaded to their cloud. For proprietary code, on-device transcription matters.
Works the same as Cursor and VS Code. JetBrains IDEs are text input fields like any other Mac app. Murmur, Wispr Flow, and Apple Dictation all paste into IntelliJ, PyCharm, WebStorm, RubyMine. The screen-context capture also picks up JetBrains windows.
Murmur's screen-context AI is the cheapest way to get Cursor-aware dictation on a Mac. Quick mode is free; the $29 licence unlocks AI mode and Silent mode.