huske 0.3.0

Always-on terminal audio recorder + local transcription (Whisper)

huske

huske — Norwegian for "to remember"

A terminal app that runs in the background, continuously records your microphone plus your computer's system audio, and transcribes the audio locally with mlx-whisper — producing a day-organized, LLM-friendly knowledge base of everything that was said on your machine throughout the day.

Point Claude Code (or any other LLM agent) at ~/huske/transcripts/ and ask it about your day.

~/huske/transcripts/
├── 2026-05-07/
│   ├── 091500_8a3f2c19_001.md
│   ├── 093000_8a3f2c19_002.md
│   └── 094500_8a3f2c19_003.md
└── README.md

Features

• Continuous capture — mic (sounddevice) + system audio (Apple ScreenCaptureKit), mixed in software, no gaps at chunk boundaries.
• No drivers, no Audio MIDI Setup — system audio comes through Apple's modern ScreenCaptureKit framework. Just grant Screen Recording permission once.
• Local transcription — mlx-whisper, default base model, runs on the Apple Silicon GPU via MLX. Audio never leaves your machine.
• Configurable chunk size — default 15 minutes, anything from 6 s to 60 min.
• Resilient — graceful stop finalizes the partial chunk; SIGKILL + restart auto-recovers orphaned audio.
• Pretty terminal UI — Rich Live panel with countdown, mic + system level meters, queue depth, last-saved transcript, rolling event log.
• LLM-ready output — every transcript is a single Markdown file with full YAML frontmatter; the directory layout is documented in ~/huske/transcripts/README.md (auto-generated).
• Optional periodic screenshots — opt in with --screenshots to also capture a JPEG of every attached display every 10 s, stored under ~/huske/screenshots/YYYY-MM-DD/<session>/HHMMSS_dN.jpg for downstream multimodal LLM use. Off by default; see Periodic screenshots.

Requirements

• macOS 13 (Ventura) or newer. Apple Silicon is the primary target.
• Python 3.11, 3.12, or 3.13.

Quickstart

# 1. Install
uv tool install huske

# 2. Validate setup (will prompt for Screen Recording permission on first run)
huske doctor

# 3. Record (Ctrl+C to stop)
huske run

# 4. Reclaim orphans from a prior crash without recording
huske recover

Other install options:

# Alternative Python tool installer:
pipx install huske

# macOS Apple Silicon with Homebrew:
brew tap tiagomoraes/huske
brew install huske

On first launch macOS will prompt you to grant Screen Recording permission to your Python interpreter — that's what ScreenCaptureKit needs to capture system audio. After approving once, it's silent forever.

For prerelease builds or exact GitHub tags, install directly from the repository:

uv tool install "git+https://github.com/tiagomoraes/huske.git@v0.1.0"

See quickstart.md for the full setup.

Update notifications

On startup, huske checks PyPI at most once every 24 hours and prints an "update available" banner with the right upgrade command for your install method (uv tool upgrade huske, pipx upgrade huske, or brew upgrade huske). The check runs in a background thread, is silent on network errors, on non-TTY stderr, and for editable installs. Disable it with:

export HUSKE_NO_UPDATE_CHECK=1

Periodic screenshots (opt-in)

huske run --screenshots enables a background thread that captures a JPEG of every attached display every 10 seconds (configurable). Screenshots are written to:

~/huske/screenshots/YYYY-MM-DD/<session_id>/HHMMSS_dN.jpg

…where dN is the display index (d1 is the main display). Filenames are timestamped so a downstream multimodal LLM can correlate each screenshot with that day's transcripts.

Capture uses macOS's built-in screencapture, so no extra dependency is needed and the same Screen Recording permission you've already granted for system audio is reused.

Flags:

huske run --screenshots                       # opt in
huske run --screenshots --screenshot-interval 30
huske run --screenshots --screenshots-root ~/another/path

Privacy. Screenshots can capture passwords, private chats, financial details, and anything else on screen. They're stored unencrypted on disk and read-accessible to any process running as your user. Treat ~/huske/screenshots/ exactly like the audio and transcript directories: never commit it, share with care, and review the Privacy and consent section before enabling.

Privacy and consent

huske is local-first: audio capture and transcription run on your machine, and the app writes transcripts to your configured filesystem path. That does not make the data low-risk. Recordings, transcripts, logs, filenames, and device metadata can contain private or legally sensitive information.

• Get consent before recording other people or regulated conversations.
• Do not commit generated audio, transcripts, logs, local configs, model caches, or screenshots containing private content.
• The --screenshots flag captures everything visible on every attached display every 10 s — including any password manager popovers, banking tabs, or private DMs that happen to be open. Leave it off unless you've consciously decided you want this in the on-disk record.
• Redact huske doctor output before sharing it publicly.
• Report security or privacy vulnerabilities privately through SECURITY.md.

Documentation

• Development — local setup, checks, and test strategy.
• Contributing — how to open issues and pull requests.
• Issue triage — labels and maintainer workflow.
• Release checklist — release preparation notes.
• Spec — what huske does and why.
• Plan — technical context and architecture.
• CLI contract — flags, exit codes.
• Transcript format contract — the LLM-consumer interface.
• Quickstart — end-to-end setup.

Community

• Follow the Code of Conduct.
• Use issue templates for bugs, features, and documentation reports.
• Use the pull request template and include the exact checks you ran.
• For help, see SUPPORT.md.

License

huske is released under the MIT License. Third-party notices are in THIRD_PARTY_NOTICES.md.