What works today and what does not
State of the code on 2026-11-07: version 0.1.0, no public release, no download. No benchmarks, no accuracy numbers: I have not measured them yet.
Exists in the code
The core
- Global push-to-talk shortcut.
- Local Whisper transcription, Portuguese language.
- Language-model formatting, instructed to write European Portuguese.
- Pasting into the active window and detecting the application name.
In preparation
Platforms
- macOS (Apple Silicon) and Windows are the targets of the automated builds.
- Linux is not supported: pasting and window detection do not exist there.
- No version has been published and the auto-updater has no key configured yet.
Undecided
Architecture
- Two ways to transcribe (a Rust engine and a Python helper) and two to format (OpenRouter or a Supabase function). One of each will remain.
- Pricing and plans: the code has a free limit and a draft paid plan; nothing is settled.
Does not exist
Not promised
- Spoken punctuation commands (“vírgula”, “ponto final”, “nova linha”): punctuation is inferred by the model.
- A fully offline mode: formatting needs the cloud. Without it you get Whisper’s text as it is.
- A rule-based AO90 guarantee: the instructions ask for European Portuguese and do not fix the spelling agreement.
- Personal dictionary, text snippets and styles: the app has screens for them, but I have not confirmed they change the output, so I do not advertise them.