openquack

Using OpenQuack with Claude Code

Claude Code (the CLI) takes long, intent-rich prompts well. Typing those prompts is the bottleneck. Speaking them through OpenQuack removes it.

This guide is for people who already use Claude Code. If you’re new to it, install via npm install -g @anthropic-ai/claude-code and run claude in any directory; the documentation lives at claude.com/claude-code.

The pairing in one sentence

Press the OpenQuack hotkey, describe what you want Claude Code to do (as long as you’d think it), release — the transcript appears at your cursor in the Claude Code prompt. Press Enter.

Why it works well

Claude Code prompts have an unusual shape: they reward detail and context. “Fix the bug” is worse than “There’s a race in the streaming-pump frame ordering — Task spawning per buffer doesn’t guarantee FIFO at the actor; the AsyncStream pump fixes it. Apply that pattern in Sources/OpenQuackApp/OpenQuackApp.swift line ~498.” The longer prompt is faster to speak than to type. Voice collapses the typing tax on detail.

OpenQuack is built for this specific shape — long-form dictation where the wait time after stop stays constant in clip length. A 30-second prompt finishes pasting ~3 seconds after you stop talking (see docs/BENCHMARKS.md). On most other dictation tools, the post-stop wait scales with your prompt length.

Setup (5 minutes)

Install OpenQuack (docs/INSTALL.md).
Grant Microphone + Accessibility when macOS asks. Accessibility is what lets the transcript paste at your cursor in the terminal; without it the transcript only goes to the clipboard.
Pick a hotkey in Settings → Shortcut that you’d want to press while your hands are on the keyboard. The default ⌃⇧Space works well in most terminals; alternatives that don’t conflict with common terminal bindings: F5, F12, ⌥⇧V.
In Claude Code, position your cursor in the prompt area, press the hotkey, speak your intent, press the hotkey again to stop. Press Enter to send.

Workflow patterns that pair well

Steering the agent. Describe the higher-level direction and constraints, then let Claude Code handle the implementation. Voice is faster than typing for this because most of what you’d say is intent, not code. Example: “Read the polish engine spec, scope down PR #2 to just the protocol surface, file it as a draft PR.”

Reviewing diffs aloud. Read what Claude Code wrote back to it as you scan. “That’s right — but rename currentPolishEngine to makePolishEngine since it constructs rather than fetches. Apply.” Voice keeps your hands free to scroll.

Running approval prompts. Claude Code asks for approval before running commands. Voice the approval if your hands are away from the keyboard. “Yes, run it.”

Dictating commit messages. OpenQuack pastes anywhere a cursor sits, including the git commit -e editor. Speak the commit body naturally; the multi-paragraph cadence translates well.

Pitfalls

Don’t dictate code blocks. Whisper transcribes prose, not syntax. If you need to drop in a code snippet, copy-paste — voice is for the surrounding intent.
Proper nouns and jargon. Add project names, file paths, and technical terms to Settings → General → Custom dictionary (one per line). Whisper biases toward those words during decoding. Without it: “Whisper Kit” might come out two words; “Claude Code” might come out “cloud code.”
First transcription is slow. Whisper has a cold start on Apple Silicon (~10-30 s for medium on M-series). Open OpenQuack before Claude Code so the model is warm when you press the hotkey.

Why not just type

You can. The point isn’t that voice is universally better — it’s that for long, detailed prompts, voice is faster end-to-end. A 40-word prompt takes ~10 seconds to speak and ~3 seconds to paste. Same prompt takes 30+ seconds to type at 90 WPM (and that’s a typing speed most people don’t sustain).

If you mostly send 5-word prompts, OpenQuack doesn’t help. If your prompts to Claude Code are paragraphs of context, it’s the right shape.

Quick voice-to-action (kickoff mode)

Sometimes you don’t want to dictate into a prompt field — you want to start a Claude Code session with what you just said. Set a timer. List a folder. Sketch a one-file prototype. Whatever Claude Code can do from a fresh shell.

Bind a second hotkey in Settings → Shortcut → Agent kickoff (e.g. ⌃⇧K). When you press it:

OpenQuack records and transcribes as normal
Instead of pasting at your cursor, it opens a new Terminal window in ~/OpenQuackAgent/ with claude already running on your transcript as the seed prompt
You see the session work in that window — approvals, output, side effects — and can continue typing or speaking into it

Your normal dictation hotkey is unaffected and keeps pasting locally. The kickoff hotkey is opt-in: the first time you bind it, OpenQuack asks for consent because the transcript now leaves your machine via Claude Code’s API call. Revoke any time by clearing the binding in Settings.

Useful when:

You’re not in Claude Code. Browsing, in Mail, in a meeting — you want a small task done without switching contexts.
The agent is the actuator, not the editor. “Move my downloaded conference talks into ~/Movies/talks/”, “Open a PR template for the bench update I keep forgetting”, “Tell me what’s listening on port 5432.”
You want a scratch workspace. ~/OpenQuackAgent/ is isolated — Claude can scaffold prototypes there without touching your real repos.

Privacy

Claude Code itself routes your prompts through Anthropic’s API under your auth. OpenQuack doesn’t change that — it just gets the prompt from your voice into the prompt field. Audio and the transcript stay on your Mac (transcribed locally via WhisperKit); nothing OpenQuack does adds a network hop in dictation mode.

The agent-kickoff hotkey (above) is the one exception: in that mode, the transcript is handed to claude, which routes through Anthropic under your existing credentials. The kickoff hotkey ships unbound and is opt-in with a consent prompt that names this destination explicitly. Dictation-mode is untouched by that consent.

The privacy gradient is yours to set per Claude Code’s settings.

This site is open source. Improve this page.