Built a Free Superwhisper Alternative Using Claude Code
Built a Free Superwhisper Alternative Using Claude Code
Superwhisper is great but it costs money and sends your audio somewhere. If you care about privacy or just want to avoid a subscription, you can build the same thing locally using Whisper and Claude Code in a weekend.
Why Local Whisper
The latency argument for cloud transcription used to be strong. Cloud models were faster. But whisper.cpp running on Apple Silicon now transcribes in near real-time. The small model handles English with surprisingly high accuracy. The medium model catches edge cases. Neither requires an internet connection.
Privacy is the bigger win. Voice input captures everything you say - half-formed thoughts, confidential client names, passwords you accidentally read aloud. Sending that to a cloud service means trusting their data handling. Running it locally means the audio never leaves your machine.
The Setup
The core is simple: capture audio from the microphone, pipe it to whisper.cpp, and inject the transcribed text at the cursor position. On macOS, you can use the accessibility API to place text in any application without clipboard hacking.
Claude Code made the implementation fast. Describe the audio capture pipeline, the Whisper integration, and the text injection mechanism. The agent writes the Swift code, handles the microphone permissions, and wires up the global hotkey. What would take a weekend of documentation reading takes an afternoon.
What You Get
A menu bar app that listens when you hold a hotkey, transcribes locally, and types the result wherever your cursor is. No subscription. No cloud. No audio leaving your machine. Latency under 500ms for most utterances.
The tradeoff is accuracy on specialized vocabulary. Cloud services fine-tune on massive datasets. Local Whisper works with what it has. For everyday dictation and voice commands to your agent, it is more than enough.
Fazm is an open source macOS AI agent. Open source on GitHub.