Releases: misty-step/cantrip
Release list
Cantrip v0.1.1
Improvements in v0.1.1
- Transcription and optional cleanup now have distinct pixel rhythms, without implying unmeasured progress. The listening waveform is unchanged.
- Successful delivery settles into a clear checkmark and stays visible before fading. Clipboard-only and cleanup-fallback outcomes keep their separate guidance; reduced motion remains supported.
- Start from a verified Linux download, then follow one attended first-dictation journey with explicit terminal, shortcut, editor, and desktop-support boundaries.
- Release archives now include focused installation, usage, configuration, desktop, and privacy guides, including retained-recording recovery and the limits of Forget.
Verified build identity
- Version:
v0.1.1 - Source:
ee69414cd520b11dead3e243928140ca078f22b3 - Target:
x86_64-unknown-linux-gnu, CPU-only; Rust1.98.1 - Runtime baseline: Ubuntu 24.04, glibc 2.39 or newer
- Archive SHA-256:
d5e906a25f6a8be70ae48e665606c60b49245bf5323f0170a240a69efef3c825 - Installation, maintenance, limitations, and verification instructions are in the archive's
INSTALLATION.md.
Cantrip v0.1.0
First public Linux release
Cantrip is local-first Linux dictation: record a take, transcribe it on your CPU, and deliver the text only when the destination can be verified.
Included
- A standalone CPU-only Linux x86-64 download. Installing it does not require a source checkout, Rust, a GPU, or CUDA.
- Local Parakeet speech recognition. Download the model explicitly once with
cantrip models pull; it is not bundled in the executable.cantrip transcribe --localdoes not use cloud recognition or cleanup. - Recording controls, a native settings window, a compact Wayland status HUD, and explicit actions for reviewing and recovering takes.
- Private local transcript history and retained stopped recordings. Dismissal does not delete a take; deleting retained audio requires an explicit Forget confirmation.
- Optional OpenAI-compatible cloud recognition and transcript cleanup. A failed configured-cloud recognition attempt gets one whole-take attempt with the installed default local model, without an implicit download or provider switch.
- Binary-only installation, explicit-backup updates, rollback, and uninstall. These operations leave configuration, models, keyring credentials, history, and startup ownership alone.
Requirements and limits
The release baseline is Ubuntu 24.04 / glibc 2.39 on Linux x86-64. Microphone dictation needs a working PipeWire session and pw-record; clipboard/paste delivery needs wl-copy.
Automatic keyboard delivery requires the documented verified Hyprland/logind desktop. Unknown focus, locking, suspension, or an interrupted session defers delivery rather than guessing where to type. Other desktops can use explicit clipboard delivery. The HUD requires Wayland layer-shell support. A successful command-line transcription check does not certify every compositor or microphone.
Installation does not enable a service, add shortcuts, or replace an existing startup owner. Read the bundled INSTALLATION.md, run cantrip doctor, and make an attended first dictation in a safe destination. Models remain a separate explicit download; cloud features remain opt-in.
This is the first published artifact. Updates retain the prior executable at an explicitly chosen backup path. Restoring an older executable does not undo data-format changes; consult the target release's notes before rolling back across future schema changes.
Verified build identity
- Version:
v0.1.0 - Source:
b837c4073f85bf7556a689b2994cecb54e970c99 - Target:
x86_64-unknown-linux-gnu, CPU-only; Rust1.98.1 - Runtime baseline: Ubuntu 24.04, glibc 2.39 or newer
- Archive SHA-256:
2e94a09d126c64a39f9ac6662ec93402cede35a42b80c307979692beda038b3e - Installation, maintenance, limitations, and verification instructions are in the archive's
INSTALLATION.md.