Longhand is a macOS dictation app. You speak, Whisper transcribes the recording on your Mac, and Claude rewrites the transcript into written language. Each dictation keeps all three: the audio, the transcript, and the rewrite.
Press ⌘R to record, ⌘R again to stop.
- macOS 26 or later, on an Apple silicon Mac
- A microphone
- Claude Code, installed and signed in. Longhand runs the rewrite through your local
claudebinary. - 2.8 GB of disk space for the transcription model
Longhand downloads Whisper large-v3-turbo from Hugging Face to ~/Library/Application Support/Longhand/Models: 1.6 GB of ggml weights and a 1.2 GB Core ML encoder. You can record while it downloads. Those recordings transcribe as soon as the model is ready.
macOS asks for microphone access the first time you record.
Speech recognition is fixed to German. To change it, edit the language string in WhisperTranscriber.runInference.
Recordings, transcripts, and rewrites stay in ~/Library/Application Support/Longhand. There's no analytics and no account.
The rewrite is the only step that sends text off your Mac. The transcript goes to the Anthropic API through your local claude CLI, under your own Claude subscription. The audio never leaves the machine.
Longhand runs claude without tools, MCP servers, slash commands, user settings, or session persistence. The transcript arrives on stdin as data, not as instructions.
xcodebuild -project Longhand.xcodeproj -scheme Longhand -configuration Release buildThe app builds to DerivedData. To build into the project folder instead, add -derivedDataPath build. No Apple Developer account is needed: the project signs to run locally.
Longhand isn't sandboxed, because it launches claude from your PATH.
Shorthand is what you take down while someone speaks. Longhand is what you turn it into.
Longhand is small on purpose, and it stays that way. Bug fixes and prompt improvements are welcome; new features usually aren't. CONTRIBUTING.md says what gets merged, how to build the project, and why there is no test suite.
Stuck on something instead? SUPPORT.md points you at the right place.
Found a security issue? Don't open an issue — report it privately. SECURITY.md describes the threat model.
Everyone taking part follows the Code of Conduct. Longhand is built with AI coding tools; AI_DISCLOSURE.md says exactly how.
Longhand is released under the MIT License.
It stands on two pieces of work, both MIT-licensed:
- whisper.cpp by the ggml authors, which runs the transcription
- Whisper by OpenAI, whose
large-v3-turboweights it downloads
Their license texts are in NOTICE.