A live transcript page and recording controls for hark, the macOS CLI that captures system audio and the microphone and transcribes them on-device.
hark-viewer shows the transcript in the browser while hark writes it, with a colour per speaker and a clock time per line, and puts Record, Stop, Pause and Mute on the same page. It also ships an agent skill, so a coding agent such as Claude Code can start a recording and answer questions about the call while it is happening.
- Records the whole computer plus your microphone through hark's Core Audio tap, with no virtual audio driver.
- Keeps your microphone on the left channel and the call on the right.
- Labels lines
Youfor the microphone andSpeaker 1..Nfor voices on the computer side. - Shows the line still being spoken in grey, when hark can stream it.
- Starts, stops, pauses and mutes from the page. No terminal window stays open.
- Warns when the capture dies, and restarts it into the same call folder.
- Files every call in its own folder, keeps the audio, and writes an accurate transcript once the call stops (its call side needs a hark newer than 0.4.3).
- Says which languages a call was in, and fixes live speaker labels after the call with
relabel. - Runs on
127.0.0.1only. Nothing leaves the machine.
git clone https://github.com/GeiserX/hark-viewer.git
cd hark-viewer
./hark-viewer work "Weekly sync" # record a call filed under "work" and open the pageNeeds macOS 14.4 or later on Apple Silicon, hark 0.4.3 or later with the Parakeet model, and the Python 3 macOS ships. Setup is in Getting started.
The full documentation is at geiserx.github.io/hark-viewer.
- Getting started: requirements, the hark models, first recording
- Configuration: every setting and its default
- Usage: every command, URL options, where calls are filed
- How it works: the server, its API, what it asks hark to record
- Troubleshooting: when the capture dies mid-call, the warnings, Restart, and how to report a bug
- Development: running the tests
- The accurate transcript: the offline pass after a call, and language detection
- As an agent skill: Claude Code, fixing speakers, getting your own voice back out
- Limits: what it cannot do yet. Recording a call needs the consent of the people on it.