Give Claude a podcast, meeting recording, or any audio/video file and get back a transcript it can actually use — chaptered, timestamped, and split into speaker turns.
amagi-listen downloads the audio if you give it a URL (same yt-dlp pipeline
as amagi-watch), transcribes it locally with faster-whisper, splits the
transcript into chapters at the longest pauses, and optionally marks
probable speaker turn breaks. Everything lands in one index.md.
Everything runs locally. There are no API keys and nothing leaves your machine except yt-dlp downloading the source, if you pass a URL.
pipx install amagi-listen
You also need these on your PATH:
Python 3.10 or newer. faster-whisper is installed as a dependency; the first transcription downloads the Whisper model you pick.
amagi-listen <file-or-url> [-o OUTDIR] [--model small] [--chapters 6] [--speakers] [--pause-threshold 1.5]
| Option | What it does |
|---|---|
-o, --outdir |
Where run folders go (default amagi_listen_out) |
--model |
faster-whisper size: tiny, base, small (default), medium, large-v3 |
--chapters |
Target number of chapters, split at the longest pauses (default 6) |
--speakers |
Mark speaker turn breaks (pause-length heuristic — see Limitations) |
--pause-threshold |
Seconds of silence that counts as a new turn (default 1.5) |
--keep-audio |
Keep the downloaded/normalized audio file |
Examples:
# A local meeting recording
amagi-listen ./standup_2026-10-19.m4a -o runs --speakers
# A podcast episode by URL
amagi-listen "https://example.com/episode-42" -o runs --chapters 8
The last line of stdout is the path to the run folder.
Each run writes <outdir>/<name>/:
index.md: the file to read. Chapters in time order, each with an extractive title (the first ~12 words of its first line — not an LLM summary) and its timestamped lines.transcript.md: the flat, unchaptered transcript.metadata.json: title, source, duration, and segment/chapter counts.
Real output from a short synthesized two-speaker test clip:
# demo_meeting
- Source: demo_meeting.wav
- Speaker turns: pause-length heuristic, not voice identification
## Chapters
### Chapter 1 — [00:00] Welcome to the AmagiLabs Weekly Sync.
**[00:00]** Speaker A: Welcome to the AmagiLabs Weekly Sync.
**[00:03]** Speaker A: Today we are covering three topics.
**[00:08]** Speaker A: First, the amagi screen launch went well.
### Chapter 2 — [00:11] Second, amagi listen is almost ready to ship.
**[00:11]** Speaker A: Second, amagi listen is almost ready to ship.(That test clip used one synthesized voice for both halves, split only by a silence gap — which is why it's one speaker label throughout. See Limitations below.)
The repo includes a SKILL.md. Copy it into your skills folder:
mkdir -p ~/.claude/skills/amagi-listen
cp SKILL.md ~/.claude/skills/amagi-listen/SKILL.md
When you paste an audio file or URL and ask what was said, Claude runs
amagi-listen and reads index.md.
--speakersis a pause-length heuristic, not voice-based diarization. It has no model of who is speaking — only that a long-enough silence probably means a new turn. Two people talking without a pause between them get mislabeled as one speaker. Treat the labels as approximate.- Chapter titles are extractive (the first words of the chapter), not generated summaries — amagi-listen doesn't call an LLM.
- Whisper on CPU runs at roughly real time or slower with
small. - URL downloads depend on yt-dlp supporting the site and the content being publicly accessible.
- amagi-watch — give Claude a video URL and get back a transcript + key frames.
- amagi-screen — give Claude a screen recording and get back a frame timeline.
- amagi-legal-review-skills
- More tools at github.com/amagilabs.
If amagi-listen saves you from replaying a two-hour recording, a star helps other people find it.
Only download recordings you have the right to download. Follow each platform's terms of service and respect copyright. amagi-listen is for personal research and analysis.
MIT. Copyright (c) 2026 AmagiLabs (Amagisolutionsgroup, Inc.). See LICENSE.
Built by AmagiLabs.
